Merge pull request 'feat(learning): chair-feedback style corrections auto-flow to writer (P1 #6)' (#344) from worktree-chair-feedback-flow into main

The chair-feedback chain was DEAD: 27 feedback rows captured, 0 ever became a
lesson the writer reads (no weekly-analysis job ever existed). The chair's own
corrections — the most authoritative signal of all, and exactly "learn from every
returned draft" — went nowhere.

decision_lessons can't carry them: it's FK-coupled to a style_corpus row (a signed
final), which a case-in-progress doesn't have. So chair STYLE feedback instead
rides the discussion_rules channel that already reaches the writer for all blocks —
the same path /training promote uses.

- db.append_global_rule: the locked read-modify-write append, extracted from web
  `_append_methodology_override` into one shared impl (G2). The web function is now
  a thin wrapper that seeds defaults; chair-feedback calls it directly.
- record_chair_feedback (MCP tool): a STYLE-category correction (style/wrong_tone/
  wrong_structure) with a lesson_extracted flows immediately to discussion_rules.
  The chair IS the gate — no separate approval (INV-LRN1 graduated gate). SUBSTANCE
  feedback (missing_content/factual_error/other) is case-specific → recorded only.

Invariants: INV-LRN1 (chair-authored style = highest authority, flows; substance
not auto-flowed), G2 (single append impl shared by promote + feedback), INV-LRN5
(style channel only).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.claude/agents/hermes-curator.md

@@ -15,7 +15,7 @@ hermes-curator.md — מקור-האמת היחיד לפרומפט של סוכן
   adapter:  deepseek_local · model: deepseek-v4-pro
   profiles: CMP=curator-cmp (רישוי 1xxx) · CMPA=curator-cmpa (היטל 8xxx + פיצויים 9xxx)
   role:     Knowledge Curator — מנתח החלטות סופיות אחרי export, מציע עדכוני skills/lessons.
-            read-only על תוכן; write רק על comments / interactions (G10).
+            read-only על תוכן; כותב comments / interactions + ממצאים מוצעים (decision_lessons, שער-יו"ר G10).
 
 placeholders זמינים: {{agentId}} {{agentName}} {{companyId}} {{companyName}} {{runId}}
   {{taskId}} {{taskTitle}} {{taskBody}} {{commentId}} {{wakeReason}} {{projectName}} {{paperclipApiUrl}}
@@ -34,13 +34,19 @@ case "$WAKE" in
       nohup .venv/bin/python ../scripts/final_${KIND}_pipeline.py --case "$CASE" \
       > "/tmp/final_${KIND}_${CASE}.log" 2>&1 &
     sleep 2
-    echo "PIPELINE_STARTED final_${KIND}_pipeline case=$CASE log=/tmp/final_${KIND}_${CASE}.log"
+    if [ "$KIND" = "learning" ]; then
+      echo "PIPELINE_STARTED_LEARNING case=$CASE log=/tmp/final_learning_${CASE}.log CONTINUE_TO_ANALYSIS"
+    else
+      echo "PIPELINE_STARTED_HALACHA final_${KIND}_pipeline case=$CASE log=/tmp/final_${KIND}_${CASE}.log"
+    fi
     ;;
   *) echo "NO_PIPELINE_WAKE" ;;
 esac
 ```
-אם הפלט הוא `PIPELINE_STARTED ...` — **זו כל המשימה**: כתוב comment קצר בעברית ("הופעל צינור <KIND> לתיק <CASE>; התוצאות יופיעו ב-/training (סגנון) או /approvals + /precedents (הלכות) תוך מספר דקות."), סגור את ה-issue (status=done), ו**סיים מיד — אל תמשיך לסעיפים שלמטה**.
+**ניתוב לפי הפלט:**
-אם הפלט הוא `NO_PIPELINE_WAKE` — המשך כרגיל לתבנית שלמטה.
+- `PIPELINE_STARTED_HALACHA ...` — **זו כל המשימה**: comment קצר בעברית ("הופעל צינור הלכות לתיק <CASE>; התוצאות יופיעו ב-/approvals + /precedents תוך מספר דקות."), סגור issue (status=done), **סיים מיד — אל תמשיך**.
+- `PIPELINE_STARTED_LEARNING ... CONTINUE_TO_ANALYSIS` — **מצב AUTO (mark-final)**: צינור-הפאנל רץ ברקע (אל תריץ `ingest_final_version` ידנית — ראה ההערה למטה). **המשך ל-§A** וזהה דפוסים משלך, אבל ב-AUTO: בצע §A.1–§A.5b בלבד, ואז **דלג על §A.6 (interaction) — אל תעיר את דפנה** (הממצאים `proposed` ונסקרים ב-/training); המשך ל-§A.7. רשומת `style_corpus` כבר קיימת (enroll רץ ראשון בצינור).
+- `NO_PIPELINE_WAKE` — יקיצת-תגובה/ידנית: המשך כרגיל ל-§A **כולל** §A.6 (interaction).
 
 > **הערה (INV-LRN4 / X16):** הצינור `final_learning_pipeline.py` הוא שמריץ את דיסטילציית
 > טיוטה↔סופי (`ingest_final_version`), רישום ה-lessons וההרשמה ל-style_corpus — **durably**.
@@ -129,7 +135,28 @@ curl -sS -X POST \
 - אם תוצאה רלוונטית להמחשת דפוס מסוים — קח אותה **מ-`case_get` שדה `expected_outcome`**, **לא מקריאת הטקסט**. אם השדה ריק או חסר ב-DB — סמן `[תוצאה: לא מאומתת]` או דלג עליה.
 - אל תפרש משפטית את ההחלטה. דפנה כבר הכריעה. תפקידך זיהוי דפוסים בלבד.
 
+## 5b. שמור את הממצאים מבנית (חובה — INV-LRN3)
+ה-comment הוא ארעי ולא-נסקר. כדי שהממצאים ייתפסו, יופיעו בטאב ״אוצֵר״ ב-/training, ויעברו
+שער-יו"ר — קרא לכלי ה-MCP **`mcp__legal-ai__record_curator_findings`** עם אותם ממצאים:
+```
+record_curator_findings(
+  case_number="<מספר התיק מ-taskTitle>",
+  findings=[
+    {"text": "<ניסוח הממצא — אותו טקסט כמו ב-comment, בלי התג>", "category": "style"},
+    {"text": "...", "category": "structure"},
+    ...
+  ]
+)
+```
+מיפוי תג→category: `[סגנון]`→`style` · `[מבנה]`→`structure` · `[לקסיקון משפטי]`→`lexicon` · `[טבלאי]`→`tabular`.
+הכלי כותב כל ממצא כ-`decision_lesson` (`source='curator'`, `review_status='proposed'`) ומדלג על כפילויות.
+אם הוא מחזיר שגיאת "לא נמצאה רשומת style_corpus" — הסופי טרם נקלט (מירוץ נדיר מול enroll שבצינור). **המתן ~30 שניות ונסה פעם נוספת**; אם עדיין נכשל — ציין זאת ב-comment והמשך (אל תיכשל).
+**זו הצעה הממתינה לאישור דפנה — לא שינוי-קול. אתה עדיין read-only על התוכן ולא נוגע ב-skills/קבצים.**
+
 ## 6. בחר interaction (חובה — רוב המקרים יש)
+> **במצב AUTO (יקיצת `PIPELINE_STARTED_LEARNING` מ-mark-final): דלג על כל §A.6 ועבור ל-§A.7.** אל תעלה interaction
+> ואל תעיר את דפנה — הממצאים כבר נרשמו כ-`proposed` (§A.5b) ונסקרים בטאב "אוצֵר" ב-/training. §A.6 חל רק על יקיצת-תגובה/ידנית.
+
 לפי הקונטקסט בחר **אחד** מ-3 הסוגים. אם **אין שום החלטה אנושית נדרשת** — דלג ישירות ל-§A.7.
 
 ### 6a. ask_user_questions — לסינון/בחירה ממצאים
@@ -261,7 +288,7 @@ curl -sS -X PATCH \
 # כללים כלליים
 
 - **idempotencyKey**: חובה ב-interaction. אם נעור פעמיים על אותו תיק — Paperclip לא יוצר כפילות.
-- **לא לעדכן** קבצים (skills/, lessons.py, DB) בעצמך. רק לכתוב comments / interactions.
+- **לא לעדכן** קבצים (skills/, lessons.py) או שכבת-הקול בעצמך. הפלט שלך: comments + interactions + רישום ממצאים כ-`decision_lesson` *מוצע* דרך `record_curator_findings` (שער-יו"ר, INV-LRN3). אתה לא מאשר ולא מטמיע — רק מציע.
 - **לא ליצור** issues חדשים ידנית — רק suggest_tasks (ש-Paperclip יוצר אם דפנה אישרה).
 - **לא להעיר** סוכנים אחרים.
 - **בעיה?** אם MCP נכשל או מסמך חסר — comment קצר עם הסיבה + סגור (status=done). אל תזייף.

.claude/agents/legal-analyst.md

@@ -166,7 +166,7 @@ tools:
 10. **הכנה ל-CREAC** — לכל סוגיה רשום:
     - כלל (Rule): הכלל המשפטי/תכנוני שיעמוד בבסיס הדיון
     - עובדות מפתח (Facts): העובדות שיופיעו בשלב היישום
-    - תקדים מבהיר (אם נדרש): רק אם הכלל דורש הבהרה
+    - תקדים מבהיר (אם נדרש): רק אם הכלל דורש הבהרה. **שקלל סמכות (X11):** כל תוצאת `search_precedent_library` נושאת `cited_by`/`verified`/`cite_count` — כמה ועדות-ערר אימצו את התקדים. העדף תקדים עם `cited_by.positive` גבוה, וסַמֵּן אותו ל-writer במפורש: *"אומץ ב-N החלטות ועדת-ערר"* (כדי שיבסס סמכות). תקדים עם `cited_by.negative>0` (אובחן/בוטל) — סמן זהירות, אל תציגו כ-good-law.
 11. **שאלות משפטיות** — 1-3 שאלות לפי הצורך (ראה שלב 4)
 12. **עמדת ועדת הערר** — שדה ריק שיו"ר הוועדה ימלא ידנית. **חובה להוסיף לכל סוגיה!** עמדה זו תשמש כהנחיה מחייבת לסוכן הכתיבה.
 

.claude/agents/legal-researcher.md

@@ -304,6 +304,15 @@ search_internal_decisions(
 
 **מינימום:** queries לקורפוס הסמכותי = מספר סוגיות מרכזיות שזוהו.
 
+#### 2ב.4ב — שקלול סמכות לפי טיפול-שיפוטי-מצטבר (X11) ⚠️
+
+כל תוצאה מ-`search_precedent_library` נושאת שדה **`cited_by`** + `verified`/`cite_count` — **כמה החלטות ועדת-ערר אחרות הסתמכו על אותו תקדים ואיך טיפלו בו**. זהו אות-סמכות אנושי-מצטבר (לא ניחוש-מודל), וחובה לשקלל אותו:
+
+- **`cited_by.positive` גבוה (אומץ/הוסבר ע"י הרבה ועדות)** → תקדים-עוגן. **טַען את הסמכות במפורש** בפלט-המחקר: *"הלכת [שם] (עע"מ X/YY) — אומצה ב-N החלטות ועדת-ערר"*, כדי שה-writer יוכל לבסס עליה. בין שני תקדימים שקולים-תוכן — העדף את בעל ה-`positive` הגבוה.
+- **`cited_by.negative` > 0 (אובחן/בוקר/בוטל)** → **דגל זהירות**. אל תציג תקדים שבוטל (`overruled`) כ"good law"; ציין את האבחנה. תקדים שאובחן שוב-ושוב אינו ראיה חזקה לסוגיה — בדוק את ההקשר.
+- **`unclassified`** → הטיפול טרם סווג; התייחס ל-`cite_count` הגולמי בלבד.
+- האות הזה **מחדד תיעדוף, לא מחליף קריאה** — עדיין קרא את ההלכה ואת ההקשר (INV-COR3/COR5: סמכות לסוגיה הספציפית, לא לפסק כולו).
+
 #### 2ב.4א — איתור החלטה ספציפית לפי שם — פרוטוקול לפני "לא בקורפוס" ⚠️
 
 שם תיק לבדו (למשל `"אגסי"`) **אינו מפתח חיפוש אמין**. ההטמעה הסמנטית והאינדקס הלקסיקלי בנויים על תוכן ההלכה/הפסקה — כך ששאילתת-שם עלולה להחזיר דווקא החלטות ש**מצטטות** את התיק, ולא את התיק עצמו. לפני שמכריזים שהחלטה אינה בקורפוס:

.claude/agents/legal-writer.md

@@ -274,7 +274,7 @@ case_update(case_number, status="drafted")
 ### שלב ג: לכל סוגיה — מבנה סילוגיסטי (CREAC) בקול דפנה
 1. **מסקנה** — פתח בתשובה (בקול "אנחנו" — ראה טבלה למטה)
 2. **כלל** — ציטוט סעיף החוק במלואו (לא תמצית). אם רלוונטי — סעיפי משנה כולם.
-3. **הרחבה** — תקדים רלוונטי אחד **בציטוט מלא** (לא תמצית). דפנה תמיד מצטטת בני 4-15 שורות עם הפניה `(פורסם בנבו)`.
+3. **הרחבה** — תקדים רלוונטי אחד **בציטוט מלא** (לא תמצית). דפנה תמיד מצטטת בני 4-15 שורות עם הפניה `(פורסם בנבו)`. **כשהמנתח סימן תקדים כ"אומץ ב-N החלטות ועדת-ערר" (אות-סמכות X11)** — חזק את ההסתמכות בלשון-סמכות ("הלכה מושרשת שאומצה בשורת החלטות"); זהו טיפול-שיפוטי-מצטבר אמיתי, לא ניחוש. **אל** תמציא מספר בעצמך — השתמש רק במה שהמנתח מסר (INV-AH, read-only מהמנתח).
 4. **יישום** — החל את הכלל על העובדות. הפרד ממצא עובדתי ממסקנה משפטית. השתמש בנתונים (מספרים, מידות, אחוזים).
 5. **אישור-לפני-דחייה (חובה)** — הצג את הטענה הטובה ביותר של הצד המפסיד: **"אכן [נקודה תקפה]... אולם [למה לא מכריע]"**. השימוש ב-"אכן" (לא "אמנם") הוא הסטנדרט.
 6. **למעלה מן הצורך** (חובה לטענות מרכזיות) — "גם אם היינו מקבלים את פרשנות העורר... התוצאה הייתה זהה". סוגר חלון לערעור.

.env.example

@@ -0,0 +1,12 @@
+# API Keys (Required to enable respective provider)
+ANTHROPIC_API_KEY="your_anthropic_api_key_here"       # Required: Format: sk-ant-api03-...
+PERPLEXITY_API_KEY="your_perplexity_api_key_here"     # Optional: Format: pplx-...
+OPENAI_API_KEY="your_openai_api_key_here"             # Optional, for OpenAI models. Format: sk-proj-...
+GOOGLE_API_KEY="your_google_api_key_here"             # Optional, for Google Gemini models.
+MISTRAL_API_KEY="your_mistral_key_here"               # Optional, for Mistral AI models.
+XAI_API_KEY="YOUR_XAI_KEY_HERE"                       # Optional, for xAI AI models.
+GROQ_API_KEY="YOUR_GROQ_KEY_HERE"                     # Optional, for Groq models.
+OPENROUTER_API_KEY="YOUR_OPENROUTER_KEY_HERE"         # Optional, for OpenRouter models.
+AZURE_OPENAI_API_KEY="your_azure_key_here"            # Optional, for Azure OpenAI models (requires endpoint in .taskmaster/config.json).
+OLLAMA_API_KEY="your_ollama_api_key_here"             # Optional: For remote Ollama servers that require authentication.
+GITHUB_API_KEY="your_github_api_key_here"             # Optional: For GitHub import/export features. Format: ghp_... or github_pat_...

.gitignore

@@ -6,7 +6,8 @@ data/backups/
 data/precedent-library/
 data/.auto-sync.log
 data/*.db
-data/checkpoints/        # X16 durable-pipeline SQLite checkpoints (runtime artifact)
+# X16 durable-pipeline SQLite checkpoints (runtime artifact)
+data/checkpoints/
 *.bak-pre-*
 mcp-server/.venv/
 __pycache__/
@@ -18,6 +19,30 @@ kiryat-yearim/
 continuation-prompt.md
 node_modules/
 data/eval/eval-report-*
-data/adapter-migration-state.json   # revert snapshot for migrate_agent_adapter.py (runtime state)
+# revert snapshot for migrate_agent_adapter.py (runtime state)
-.claude/agents/.generated/          # frontmatter-stripped instruction copies for content_arg adapters (generated)
+data/adapter-migration-state.json
+# frontmatter-stripped instruction copies for content_arg adapters (generated)
+.claude/agents/.generated/
 .claude/worktrees/
+
+# TaskMaster backups (runtime)
+.taskmaster/tasks/tasks.json.bak.*
+
+# Build artifacts
+.design-build/
+
+# Temp files
+.interaction_tmp.json
+
+# Runtime eval/ab-test data
+data/ab_halacha_*.json
+data/ab_run_*.log
+data/x11_treatment_run_*.log
+
+# Runtime data directories
+data/audit/
+data/bulletins/
+data/digests/
+data/internal-decisions/
+data/learning/
+data/logs/

.taskmaster/templates/example_prd.txt

@@ -0,0 +1,47 @@
+<context>
+# Overview  
+[Provide a high-level overview of your product here. Explain what problem it solves, who it's for, and why it's valuable.]
+
+# Core Features  
+[List and describe the main features of your product. For each feature, include:
+- What it does
+- Why it's important
+- How it works at a high level]
+
+# User Experience  
+[Describe the user journey and experience. Include:
+- User personas
+- Key user flows
+- UI/UX considerations]
+</context>
+<PRD>
+# Technical Architecture  
+[Outline the technical implementation details:
+- System components
+- Data models
+- APIs and integrations
+- Infrastructure requirements]
+
+# Development Roadmap  
+[Break down the development process into phases:
+- MVP requirements
+- Future enhancements
+- Do not think about timelines whatsoever -- all that matters is scope and detailing exactly what needs to be build in each phase so it can later be cut up into tasks]
+
+# Logical Dependency Chain
+[Define the logical order of development:
+- Which features need to be built first (foundation)
+- Getting as quickly as possible to something usable/visible front end that works
+- Properly pacing and scoping each feature so it is atomic but can also be built upon and improved as development approaches]
+
+# Risks and Mitigations  
+[Identify potential risks and how they'll be addressed:
+- Technical challenges
+- Figuring out the MVP that we can build upon
+- Resource constraints]
+
+# Appendix  
+[Include any additional information:
+- Research findings
+- Technical specifications]
+</PRD>

.taskmaster/templates/example_prd_rpg.txt

@@ -0,0 +1,511 @@
+<rpg-method>
+# Repository Planning Graph (RPG) Method - PRD Template
+
+This template teaches you (AI or human) how to create structured, dependency-aware PRDs using the RPG methodology from Microsoft Research. The key insight: separate WHAT (functional) from HOW (structural), then connect them with explicit dependencies.
+
+## Core Principles
+
+1. **Dual-Semantics**: Think functional (capabilities) AND structural (code organization) separately, then map them
+2. **Explicit Dependencies**: Never assume - always state what depends on what
+3. **Topological Order**: Build foundation first, then layers on top
+4. **Progressive Refinement**: Start broad, refine iteratively
+
+## How to Use This Template
+
+- Follow the instructions in each `<instruction>` block
+- Look at `<example>` blocks to see good vs bad patterns
+- Fill in the content sections with your project details
+- The AI reading this will learn the RPG method by following along
+- Task Master will parse the resulting PRD into dependency-aware tasks
+
+## Recommended Tools for Creating PRDs
+
+When using this template to **create** a PRD (not parse it), use **code-context-aware AI assistants** for best results:
+
+**Why?** The AI needs to understand your existing codebase to make good architectural decisions about modules, dependencies, and integration points.
+
+**Recommended tools:**
+- **Claude Code** (claude-code CLI) - Best for structured reasoning and large contexts
+- **Cursor/Windsurf** - IDE integration with full codebase context
+- **Gemini CLI** (gemini-cli) - Massive context window for large codebases
+- **Codex/Grok CLI** - Strong code generation with context awareness
+
+**Note:** Once your PRD is created, `task-master parse-prd` works with any configured AI model - it just needs to read the PRD text itself, not your codebase.
+</rpg-method>
+
+---
+
+<overview>
+<instruction>
+Start with the problem, not the solution. Be specific about:
+- What pain point exists?
+- Who experiences it?
+- Why existing solutions don't work?
+- What success looks like (measurable outcomes)?
+
+Keep this section focused - don't jump into implementation details yet.
+</instruction>
+
+## Problem Statement
+[Describe the core problem. Be concrete about user pain points.]
+
+## Target Users
+[Define personas, their workflows, and what they're trying to achieve.]
+
+## Success Metrics
+[Quantifiable outcomes. Examples: "80% task completion via autopilot", "< 5% manual intervention rate"]
+
+</overview>
+
+---
+
+<functional-decomposition>
+<instruction>
+Now think about CAPABILITIES (what the system DOES), not code structure yet.
+
+Step 1: Identify high-level capability domains
+- Think: "What major things does this system do?"
+- Examples: Data Management, Core Processing, Presentation Layer
+
+Step 2: For each capability, enumerate specific features
+- Use explore-exploit strategy:
+  * Exploit: What features are REQUIRED for core value?
+  * Explore: What features make this domain COMPLETE?
+
+Step 3: For each feature, define:
+- Description: What it does in one sentence
+- Inputs: What data/context it needs
+- Outputs: What it produces/returns
+- Behavior: Key logic or transformations
+
+<example type="good">
+Capability: Data Validation
+  Feature: Schema validation
+    - Description: Validate JSON payloads against defined schemas
+    - Inputs: JSON object, schema definition
+    - Outputs: Validation result (pass/fail) + error details
+    - Behavior: Iterate fields, check types, enforce constraints
+
+  Feature: Business rule validation
+    - Description: Apply domain-specific validation rules
+    - Inputs: Validated data object, rule set
+    - Outputs: Boolean + list of violated rules
+    - Behavior: Execute rules sequentially, short-circuit on failure
+</example>
+
+<example type="bad">
+Capability: validation.js
+  (Problem: This is a FILE, not a CAPABILITY. Mixing structure into functional thinking.)
+
+Capability: Validation
+  Feature: Make sure data is good
+  (Problem: Too vague. No inputs/outputs. Not actionable.)
+</example>
+</instruction>
+
+## Capability Tree
+
+### Capability: [Name]
+[Brief description of what this capability domain covers]
+
+#### Feature: [Name]
+- **Description**: [One sentence]
+- **Inputs**: [What it needs]
+- **Outputs**: [What it produces]
+- **Behavior**: [Key logic]
+
+#### Feature: [Name]
+- **Description**:
+- **Inputs**:
+- **Outputs**:
+- **Behavior**:
+
+### Capability: [Name]
+...
+
+</functional-decomposition>
+
+---
+
+<structural-decomposition>
+<instruction>
+NOW think about code organization. Map capabilities to actual file/folder structure.
+
+Rules:
+1. Each capability maps to a module (folder or file)
+2. Features within a capability map to functions/classes
+3. Use clear module boundaries - each module has ONE responsibility
+4. Define what each module exports (public interface)
+
+The goal: Create a clear mapping between "what it does" (functional) and "where it lives" (structural).
+
+<example type="good">
+Capability: Data Validation
+  → Maps to: src/validation/
+    ├── schema-validator.js      (Schema validation feature)
+    ├── rule-validator.js         (Business rule validation feature)
+    └── index.js                  (Public exports)
+
+Exports:
+  - validateSchema(data, schema)
+  - validateRules(data, rules)
+</example>
+
+<example type="bad">
+Capability: Data Validation
+  → Maps to: src/utils.js
+  (Problem: "utils" is not a clear module boundary. Where do I find validation logic?)
+
+Capability: Data Validation
+  → Maps to: src/validation/everything.js
+  (Problem: One giant file. Features should map to separate files for maintainability.)
+</example>
+</instruction>
+
+## Repository Structure
+
+```
+project-root/
+├── src/
+│   ├── [module-name]/       # Maps to: [Capability Name]
+│   │   ├── [file].js        # Maps to: [Feature Name]
+│   │   └── index.js         # Public exports
+│   └── [module-name]/
+├── tests/
+└── docs/
+```
+
+## Module Definitions
+
+### Module: [Name]
+- **Maps to capability**: [Capability from functional decomposition]
+- **Responsibility**: [Single clear purpose]
+- **File structure**:
+  ```
+  module-name/
+  ├── feature1.js
+  ├── feature2.js
+  └── index.js
+  ```
+- **Exports**:
+  - `functionName()` - [what it does]
+  - `ClassName` - [what it does]
+
+</structural-decomposition>
+
+---
+
+<dependency-graph>
+<instruction>
+This is THE CRITICAL SECTION for Task Master parsing.
+
+Define explicit dependencies between modules. This creates the topological order for task execution.
+
+Rules:
+1. List modules in dependency order (foundation first)
+2. For each module, state what it depends on
+3. Foundation modules should have NO dependencies
+4. Every non-foundation module should depend on at least one other module
+5. Think: "What must EXIST before I can build this module?"
+
+<example type="good">
+Foundation Layer (no dependencies):
+  - error-handling: No dependencies
+  - config-manager: No dependencies
+  - base-types: No dependencies
+
+Data Layer:
+  - schema-validator: Depends on [base-types, error-handling]
+  - data-ingestion: Depends on [schema-validator, config-manager]
+
+Core Layer:
+  - algorithm-engine: Depends on [base-types, error-handling]
+  - pipeline-orchestrator: Depends on [algorithm-engine, data-ingestion]
+</example>
+
+<example type="bad">
+- validation: Depends on API
+- API: Depends on validation
+(Problem: Circular dependency. This will cause build/runtime issues.)
+
+- user-auth: Depends on everything
+(Problem: Too many dependencies. Should be more focused.)
+</example>
+</instruction>
+
+## Dependency Chain
+
+### Foundation Layer (Phase 0)
+No dependencies - these are built first.
+
+- **[Module Name]**: [What it provides]
+- **[Module Name]**: [What it provides]
+
+### [Layer Name] (Phase 1)
+- **[Module Name]**: Depends on [[module-from-phase-0], [module-from-phase-0]]
+- **[Module Name]**: Depends on [[module-from-phase-0]]
+
+### [Layer Name] (Phase 2)
+- **[Module Name]**: Depends on [[module-from-phase-1], [module-from-foundation]]
+
+[Continue building up layers...]
+
+</dependency-graph>
+
+---
+
+<implementation-roadmap>
+<instruction>
+Turn the dependency graph into concrete development phases.
+
+Each phase should:
+1. Have clear entry criteria (what must exist before starting)
+2. Contain tasks that can be parallelized (no inter-dependencies within phase)
+3. Have clear exit criteria (how do we know phase is complete?)
+4. Build toward something USABLE (not just infrastructure)
+
+Phase ordering follows topological sort of dependency graph.
+
+<example type="good">
+Phase 0: Foundation
+  Entry: Clean repository
+  Tasks:
+    - Implement error handling utilities
+    - Create base type definitions
+    - Setup configuration system
+  Exit: Other modules can import foundation without errors
+
+Phase 1: Data Layer
+  Entry: Phase 0 complete
+  Tasks:
+    - Implement schema validator (uses: base types, error handling)
+    - Build data ingestion pipeline (uses: validator, config)
+  Exit: End-to-end data flow from input to validated output
+</example>
+
+<example type="bad">
+Phase 1: Build Everything
+  Tasks:
+    - API
+    - Database
+    - UI
+    - Tests
+  (Problem: No clear focus. Too broad. Dependencies not considered.)
+</example>
+</instruction>
+
+## Development Phases
+
+### Phase 0: [Foundation Name]
+**Goal**: [What foundational capability this establishes]
+
+**Entry Criteria**: [What must be true before starting]
+
+**Tasks**:
+- [ ] [Task name] (depends on: [none or list])
+  - Acceptance criteria: [How we know it's done]
+  - Test strategy: [What tests prove it works]
+
+- [ ] [Task name] (depends on: [none or list])
+
+**Exit Criteria**: [Observable outcome that proves phase complete]
+
+**Delivers**: [What can users/developers do after this phase?]
+
+---
+
+### Phase 1: [Layer Name]
+**Goal**:
+
+**Entry Criteria**: Phase 0 complete
+
+**Tasks**:
+- [ ] [Task name] (depends on: [[tasks-from-phase-0]])
+- [ ] [Task name] (depends on: [[tasks-from-phase-0]])
+
+**Exit Criteria**:
+
+**Delivers**:
+
+---
+
+[Continue with more phases...]
+
+</implementation-roadmap>
+
+---
+
+<test-strategy>
+<instruction>
+Define how testing will be integrated throughout development (TDD approach).
+
+Specify:
+1. Test pyramid ratios (unit vs integration vs e2e)
+2. Coverage requirements
+3. Critical test scenarios
+4. Test generation guidelines for Surgical Test Generator
+
+This section guides the AI when generating tests during the RED phase of TDD.
+
+<example type="good">
+Critical Test Scenarios for Data Validation module:
+  - Happy path: Valid data passes all checks
+  - Edge cases: Empty strings, null values, boundary numbers
+  - Error cases: Invalid types, missing required fields
+  - Integration: Validator works with ingestion pipeline
+</example>
+</instruction>
+
+## Test Pyramid
+
+```
+        /\
+       /E2E\       ← [X]% (End-to-end, slow, comprehensive)
+      /------\
+     /Integration\ ← [Y]% (Module interactions)
+    /------------\
+   /  Unit Tests  \ ← [Z]% (Fast, isolated, deterministic)
+  /----------------\
+```
+
+## Coverage Requirements
+- Line coverage: [X]% minimum
+- Branch coverage: [X]% minimum
+- Function coverage: [X]% minimum
+- Statement coverage: [X]% minimum
+
+## Critical Test Scenarios
+
+### [Module/Feature Name]
+**Happy path**:
+- [Scenario description]
+- Expected: [What should happen]
+
+**Edge cases**:
+- [Scenario description]
+- Expected: [What should happen]
+
+**Error cases**:
+- [Scenario description]
+- Expected: [How system handles failure]
+
+**Integration points**:
+- [What interactions to test]
+- Expected: [End-to-end behavior]
+
+## Test Generation Guidelines
+[Specific instructions for Surgical Test Generator about what to focus on, what patterns to follow, project-specific test conventions]
+
+</test-strategy>
+
+---
+
+<architecture>
+<instruction>
+Describe technical architecture, data models, and key design decisions.
+
+Keep this section AFTER functional/structural decomposition - implementation details come after understanding structure.
+</instruction>
+
+## System Components
+[Major architectural pieces and their responsibilities]
+
+## Data Models
+[Core data structures, schemas, database design]
+
+## Technology Stack
+[Languages, frameworks, key libraries]
+
+**Decision: [Technology/Pattern]**
+- **Rationale**: [Why chosen]
+- **Trade-offs**: [What we're giving up]
+- **Alternatives considered**: [What else we looked at]
+
+</architecture>
+
+---
+
+<risks>
+<instruction>
+Identify risks that could derail development and how to mitigate them.
+
+Categories:
+- Technical risks (complexity, unknowns)
+- Dependency risks (blocking issues)
+- Scope risks (creep, underestimation)
+</instruction>
+
+## Technical Risks
+**Risk**: [Description]
+- **Impact**: [High/Medium/Low - effect on project]
+- **Likelihood**: [High/Medium/Low]
+- **Mitigation**: [How to address]
+- **Fallback**: [Plan B if mitigation fails]
+
+## Dependency Risks
+[External dependencies, blocking issues]
+
+## Scope Risks
+[Scope creep, underestimation, unclear requirements]
+
+</risks>
+
+---
+
+<appendix>
+## References
+[Papers, documentation, similar systems]
+
+## Glossary
+[Domain-specific terms]
+
+## Open Questions
+[Things to resolve during development]
+</appendix>
+
+---
+
+<task-master-integration>
+# How Task Master Uses This PRD
+
+When you run `task-master parse-prd <file>.txt`, the parser:
+
+1. **Extracts capabilities** → Main tasks
+   - Each `### Capability:` becomes a top-level task
+
+2. **Extracts features** → Subtasks
+   - Each `#### Feature:` becomes a subtask under its capability
+
+3. **Parses dependencies** → Task dependencies
+   - `Depends on: [X, Y]` sets task.dependencies = ["X", "Y"]
+
+4. **Orders by phases** → Task priorities
+   - Phase 0 tasks = highest priority
+   - Phase N tasks = lower priority, properly sequenced
+
+5. **Uses test strategy** → Test generation context
+   - Feeds test scenarios to Surgical Test Generator during implementation
+
+**Result**: A dependency-aware task graph that can be executed in topological order.
+
+## Why RPG Structure Matters
+
+Traditional flat PRDs lead to:
+- ❌ Unclear task dependencies
+- ❌ Arbitrary task ordering
+- ❌ Circular dependencies discovered late
+- ❌ Poorly scoped tasks
+
+RPG-structured PRDs provide:
+- ✅ Explicit dependency chains
+- ✅ Topological execution order
+- ✅ Clear module boundaries
+- ✅ Validated task graph before implementation
+
+## Tips for Best Results
+
+1. **Spend time on dependency graph** - This is the most valuable section for Task Master
+2. **Keep features atomic** - Each feature should be independently testable
+3. **Progressive refinement** - Start broad, use `task-master expand` to break down complex tasks
+4. **Use research mode** - `task-master parse-prd --research` leverages AI for better task generation
+</task-master-integration>

data/halacha_night_check.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# One-shot morning verdict for the halacha night drain (scheduled 2026-06-15 04:30 UTC
+# via chaim's crontab; throwaway — lives under data/, not tracked in scripts/).
+# Captures whether last night's run (with the PR #251 fix: durable rate-limit
+# detection + 05:00–07:00 catch-up window) actually drained the backlog.
+# Baseline at install time (2026-06-14 13:xx IDT): pending=96 done=248 halachot=4099.
+set -u
+export HOME=/home/chaim
+REPO=/home/chaim/legal-ai
+PY="$REPO/mcp-server/.venv/bin/python"
+OUT="$REPO/data/logs/halacha_night_report_$(TZ=Asia/Jerusalem date +%Y%m%d).md"
+SUP_LOG=/home/chaim/.pm2/logs/legal-halacha-supervisor-out.log
+DRAIN_ERR=/home/chaim/.pm2/logs/legal-halacha-drain-error.log
+mkdir -p "$REPO/data/logs"
+
+{
+  echo "# דוח-בוקר: ריצת-הלכות הלילה — $(TZ=Asia/Jerusalem date '+%Y-%m-%d %H:%M %Z')"
+  echo
+  echo "בסיס-השוואה (אתמול 13:xx IDT): pending=96 · done=248 · halachot=4099"
+  echo
+  echo '## מצב נוכחי (supervisor status)'
+  echo '```'
+  cd "$REPO" && "$PY" scripts/halacha_drain_supervisor.py status 2>&1
+  echo '```'
+  echo
+  echo '## פעולות המתזמר ב-12 השעות האחרונות (modes/actions)'
+  echo '```'
+  grep -E 'מצב:|פעולה:|catch-up|rate-limit|נעצר' "$SUP_LOG" 2>/dev/null | tail -40
+  echo '```'
+  echo
+  echo '## אותות rate-limit בלוג-הדריינר (24ש אחרונות בלוג)'
+  echo '```'
+  echo "429 hits (tail 4000):  $(tail -4000 "$DRAIN_ERR" 2>/dev/null | grep -c '429')"
+  echo "session-limit msgs:    $(tail -4000 "$DRAIN_ERR" 2>/dev/null | grep -c 'hit your session limit')"
+  echo "extraction_failed:     $(tail -4000 "$DRAIN_ERR" 2>/dev/null | grep -c 'extraction_failed')"
+  echo "hold-stopped (fix A):  $(grep -c 'hold-stopped' "$SUP_LOG" 2>/dev/null)"
+  echo "catch-up opened (B):   $(grep -c 'catch-up בוקר' "$SUP_LOG" 2>/dev/null)"
+  echo '```'
+  echo
+  echo "_(נוצר ע\"י data/halacha_night_check.sh; ניתן למחוק את שורת ה-crontab של 15.6.)_"
+} > "$OUT" 2>&1
+
+echo "report written: $OUT"

docs/precedent-corpus-redesign/01-claude-architecture-data-audit.md

@@ -0,0 +1,170 @@
+# ממצאי ביקורת — ארכיטקטורת קורפוס־הפסיקה + מצב הדאטה בפועל
+
+> **מקור:** Claude (Opus 4.8) · **תאריך:** 2026-06-20 · **קונטקסט:** חקירה לקראת תכנון־מחדש של קורפוס־הפסיקה.
+> מסמך זה הוא **אחד מכמה** קלטי־סוכנים שחיים אוסף; ייעודו להזין את שלב־הסינתזה. אינו תכנית — הוא **אבחון**.
+>
+> **שאלת־המוצא של חיים:** "הקורפוס נבנה מראש לא נכון, אני כל הזמן מתעסק בתיקונים. האם כדאי ליצור מחדש את קורפוס־הפסיקה ולהתחיל דף נקי?"
+>
+> **המודל הרצוי (כפי שחיים תיאר אותו):** מאגר פסקי־דין והחלטות ועדות־ערר; חוקר־התקדימים מזהה בשלב ניתוח־הערר פס"ד/החלטות שדנו במקרה דומה או הלכה דומה; הסוכן־הכותב משייך ומזכיר אותם בפרק הדיון וההכרעה בסגנון דפנה. שלושה מקורות־הזנה: (1) החלטות דפנה עצמה, (2) ועדות־ערר אחרות שמצטטות פסיקה, (3) פס"ד עליון/מחוזי.
+
+---
+
+## 0. תקציר־מנהלים (TL;DR)
+
+**מה ש"בנוי לא נכון" אינו הסכמה — היא שכבת־הביצוע.** הסכמה כבר תואמת בדיוק את המודל שחיים תיאר: טבלה אחת (`case_law`) ששלושת המקורות נכנסים אליה דרך `source_kind`/`source_type`; שלושת ה־`search_*` הם שלושה *מסננים* על אותו מקור, לא שלושה מאגרים מקבילים (G2 מקוים ברמת־הסכמה). **רֵבילד של הסכמה ייצר בדיוק את אותה סכמה** — ולכן אינו פותר דבר, ומסכן ב־second-system syndrome.
+
+מה שכן מחולל את "התיקונים האינסופיים" — שלושה כשלי־ביצוע מדידים:
+1. **חוזה־קליטה רופף** → 66% מהפסיקה בלי `practice_area`, 31 רשומות ריקות, אכיפת־שלמות (INV-DM1) מופרת בפועל.
+2. **צינור הלכות→קנוני מייצר רעש** → 5,472 קנוני, מתוכם 5,456 סינגלטונים, **0 published** → השכבה שאמורה להזין את הכותב (INV-G10) **אינרטית לגמרי**.
+3. **כפילות `style_corpus`** → 55 החלטות דפנה חיות בשני נתיבי־אחזור.
+
+**המלצה:** לא לשרוף את הסכמה. כן לבצע "איפוס שכבות־נגזרות" צר (truncate ל-chunks+halachot+canonical והרצה־מחדש), **אבל רק אחרי תיקון החוזה והסף** — אחרת מחזירים את אותו בלגן (G1: תיקון במקור, לא בקריאה). מסמכי־המקור (363 רשומות, 332 עם full_text) נשמרים; הם יקרים ו/או ניתנים לקליטה־מחדש מ־PDF.
+
+---
+
+## 1. מצב הדאטה בפועל (שאילתות חיות מול legal_ai @ localhost:5433, 2026-06-20)
+
+```
+┌─────────────────────────────────────┬─────────┬──────────────────────────────────┐
+│ Metric                              │ Count   │ Reading                          │
+├─────────────────────────────────────┼─────────┼──────────────────────────────────┤
+│ case_law (total precedents)         │   363   │ קטן — re-ingestable              │
+│   • external_upload (court rulings) │   240   │ מקור (3) — עליון/מחוזי           │
+│   • internal_committee (ועדות ערר)  │    92   │ מקור (1)+(2)                     │
+│   • (יתר — ללא source_kind מובהק)   │    31   │ = הרשומות הריקות (ראה למטה)      │
+│ style_corpus (החלטות דפנה)          │    55   │ כפילות עם internal_committee     │
+│ precedent_chunks                    │ 11,904  │ נגזר — מתחדש מ-full_text         │
+│ halachot (total)                    │  5,489  │ נגזר                             │
+│   • approved/published              │  1,352  │ 25% בלבד                         │
+│   • pending_review (backlog ידני)   │  2,402  │ 44% — צוואר־בקבוק                │
+│ canonical_halachot (V41)            │  5,472  │ כמעט 1:1 עם halachot ⚠️           │
+│   • singletons (instance_count=1)   │  5,456  │ דה־דופ כמעט לא קרה                │
+│   • merged (instance_count>=2)      │    16   │ 0.3% מיזוג                       │
+│   • published (מגיע לסוכן הכותב)    │     0   │ ⚠️ השכבה אינרטית לחלוטין          │
+│ case_law w/o practice_area          │   240   │ 66% — חופף-בדיוק לפסיקה החיצונית │
+│ case_law missing summary            │    27   │                                  │
+│ case_law w/ 0 chunks / no full_text │    31   │ רשומות שבורות/ריקות              │
+│ distinct practice_area              │     4   │ rishuy/betterment/197/(ריק)      │
+└─────────────────────────────────────┴─────────┴──────────────────────────────────┘
+
+practice_area breakdown:
+  (ריק)            240   ← כל הפסיקה החיצונית ללא סיווג
+  rishuy_uvniya     70
+  betterment_levy   50
+  compensation_197   3
+```
+
+**קריאות מפתח:**
+- **240 = 240:** מספר הפסיקה־החיצונית שווה־בדיוק למספר חסרי־`practice_area`. כלומר אף פס"ד חיצוני לא סווג לתחום — סינון לפי תחום באחזור פשוט לא עובד עליהם.
+- **5,456 / 5,472 סינגלטונים:** מנוע הקנוניזציה (V41) רץ אך לא מאחד. סף 0.85 כנראה הדוק מדי, או שהחילוץ מנסח כל הלכה ייחודית מספיק כדי לא להתלכד.
+- **0 published canonical:** לפי INV-G10 רק קנוני `published` מגיע לכתיבה. אפס. **כל מנגנון V41 כרגע מנותק מהכתיבה בפועל.**
+- **2,402 pending_review:** צוואר־הבקבוק הוא אישור־אנושי ידני, לא טכנולוגיה.
+
+---
+
+## 2. מפת הארכיטקטורה — קורפוס אחד, שלושה מסננים
+
+**פיזית: טבלה אחת.** כל שלושת המקורות מתכנסים ל-`case_law`, מתחתיה `precedent_chunks` (FK) ו-`halachot` (FK), ומעל ה-`halachot` שכבת `canonical_halachot` (V41).
+
+```
+        canonical_halachot   (עקרונות מאוחדים — V41; כיום אינרטי, 0 published)
+              ▲ 1:many
+         halachot            (מופע-להלכה per precedent; review_status gate)
+              ▲ FK
+         case_law            (רישום מרכזי — court rulings + ועדות-ערר)
+         ├─ source_kind='external_upload'   → פס"ד עליון/מחוזי  [מקור 3]
+         └─ source_kind='internal_committee'→ ועדות-ערר + דפנה   [מקור 1+2]
+              ▼ FK
+         precedent_chunks    (chunks + embedding vector(1024))
+
+  בנפרד:
+         style_corpus        (55 החלטות דפנה — נתיב-אחזור מקביל ל-search_decisions)
+         document_chunks     (מסמכי-תיק + style; FK→documents→cases)
+```
+
+**שלושת נתיבי־האחזור (לא שלושה מאגרים — שלושה scopes):**
+
+```
+┌──────────────────────────────┬─────────────────────────┬────────────────────────────┐
+│ Tool                         │ Table / filter          │ Purpose                    │
+├──────────────────────────────┼─────────────────────────┼────────────────────────────┤
+│ search_decisions             │ document_chunks         │ סגנון/קול: החלטות דפנה     │
+│                              │ (scoped case/area)      │ + מסמכי-תיק                │
+│ search_precedent_library     │ case_law + chunks/halach │ source_kind=external_upload│
+│                              │ ot, source_kind filter  │ → פסיקה חיצונית            │
+│ search_internal_decisions    │ אותן פונקציות DB,        │ source_kind=internal_      │
+│                              │ source_kind אחר         │ committee → ועדות-ערר      │
+└──────────────────────────────┴─────────────────────────┴────────────────────────────┘
+```
+
+שתי האחרונות קוראות **לאותן פונקציות DB** (`search_precedent_library_semantic`/`_lexical`) עם `source_kind` שונה. זו הפרדה־בשאילתה, לא קוד מקביל.
+
+---
+
+## 3. שלושת המקורות של חיים → איפה הם נופלים היום
+
+```
+┌────────────────────────────────────┬───────────────────────────────┬─────────────────────────┐
+│ Source (חיים)                      │ Stored as                     │ Retrieval               │
+├────────────────────────────────────┼───────────────────────────────┼─────────────────────────┤
+│ (1) החלטות דפנה עצמה               │ style_corpus  +  (מהוגר ל-)   │ search_decisions  +     │
+│                                    │ case_law internal_committee   │ search_internal_decisions│
+│                                    │ chair_name='דפנה תמיר'        │ ← כפילות / נתיב-כפול     │
+│ (2) ועדות-ערר אחרות                │ case_law internal_committee   │ search_internal_decisions│
+│                                    │ chair_name=<אחר>, district     │                         │
+│ (3) פס"ד עליון/מחוזי               │ case_law external_upload      │ search_precedent_library │
+│                                    │ source_type='court_ruling'    │                         │
+└────────────────────────────────────┴───────────────────────────────┴─────────────────────────┘
+```
+
+**מסקנה:** המודל המנטלי של חיים **כבר ממומש בסכמה**. אין צורך להמציא מבנה חדש — צריך לאכוף את המבנה הקיים בקליטה, ולחבר את שכבת־הקנוני לכתיבה.
+
+---
+
+## 4. הדיאגנוזה — מה "בנוי לא נכון" (3 מחוללי־כאב)
+
+### 4.1 חוזה־קליטה רופף (root cause #1)
+- 66% מהפסיקה ללא `practice_area`; 27 ללא summary; 31 ללא full_text/chunks.
+- אין אכיפה ש"`searchable=false` עד שהמטא שלם" → INV-DM1 מופר בפועל.
+- **כל העלאה מוסיפה חוב** במקום רשומה שלמה. זה המקור לתיקונים החוזרים.
+
+### 4.2 צינור הלכות→קנוני מייצר רעש, לא ערך (root cause #2)
+- 5,456/5,472 סינגלטונים → דה־דופ לא עובד (סף 0.85? ניסוח־חילוץ?).
+- **0 published** → השכבה שאמורה להזין את הכותב (INV-G10) מנותקת.
+- 2,402 בתור־אישור־ידני → הצינור מייצר מהר יותר ממה שאדם מאשר.
+- **זה בולע את רוב זמן־התחזוקה.**
+
+### 4.3 כפילות style_corpus (root cause #3)
+- 55 החלטות דפנה בשני מקומות + שני נתיבי־אחזור.
+- צריך מקור־אמת אחד: או `style_corpus` SoT וה-`case_law` נגזר, או הפוך.
+
+---
+
+## 5. ההמלצה — לא רֵבילד־סכמה; "איפוס שכבות־נגזרות" אחרי תיקון־חוזה
+
+**אסור** לשרוף ולהעלות־מחדש את `case_law` — תבנה אותה סכמה ותאבד מטא־דאטה ידני. **כן** הגיוני רֵבילד צר של ה**נגזר**, אבל בסדר הזה (G1 — מקור לפני תסמין):
+
+1. **קודם החוזה:** אכוף ב-`*_upload` שדות־חובה (practice_area, summary, full_text); כשל → `searchable=false`. נקה/מחק את 31 הריקות.
+2. **תקן את הקנוניזציה:** כוונן סף 0.85, הגדר מתי `published`, ובדוק ניסוח־החילוץ. בלי זה אין טעם להריץ מחדש.
+3. **רק אז** re-derive מהמקור הקיים: chunks → halachot → canonical.
+4. **הכרע style_corpus:** מקור־אמת אחד.
+
+**מתי רֵבילד־מלא כן מוצדק:** רק אם יתגלה שמסמכי־המקור עצמם (PDF/full_text של 363) פגומים/חסרים. המספרים *לא* מראים זאת (332/363 עם full_text תקין).
+
+---
+
+## 6. שאלות פתוחות לשלב־הסינתזה
+- **למה 0 published בקנוני?** האם זה סף, ניסוח־חילוץ, או שפשוט אף אחד לא אישר? (קריטי — קובע אם V41 שמיש בכלל.)
+- **style_corpus מול case_law:** מי SoT? (משפיע על search_decisions מול search_internal_decisions.)
+- **practice_area לפסיקה חיצונית:** לחלץ אוטומטית בקליטה, או להשאיר ידני?
+- **תור־האישור (2,402):** האם המנגנון (פאנל/active-learning, #133) יכול לסגור את הפער, או שצריך לחתוך את קצב־החילוץ?
+
+---
+
+## 7. נספח — קבצים מרכזיים (file:line)
+- סכמה: `mcp-server/src/legal_mcp/services/db.py` (case_law, precedent_chunks, halachot, canonical_halachot)
+- קליטה חיצונית: `mcp-server/src/legal_mcp/tools/precedent_library.py` (`precedent_library_upload`)
+- קליטה פנימית: `mcp-server/src/legal_mcp/tools/internal_decisions.py` (`internal_decision_upload`)
+- הגירת style→case_law: `mcp-server/src/legal_mcp/services/internal_decisions.py` (`migrate_from_style_corpus`, chair/district hardcoded)
+- אחזור: `mcp-server/src/legal_mcp/tools/search.py`, `services/hybrid_search.py`, `services/db.py` (`search_precedent_library_semantic`/`_lexical`)
+- ספ: `docs/spec/02-data-model.md` (INV-DM1–7), `docs/spec/03-retrieval.md` (INV-RET1–5), `docs/spec/00-constitution.md` (G2/G10)

docs/precedent-corpus-redesign/06-extractor-generosity-committee-application.md

@@ -0,0 +1,125 @@
+# 06 — נדיבות-המחלץ: `application` בהחלטות-ועדה (כימות חוצה-קורפוס)
+
+> **קלט-נתונים** ליוזמה, נמדד חי על **5,489 רשומות `halachot`** (כל הקורפוס, 2026-06-20).
+> נולד משאלת-חיים: "8508-03-24 מפיק 71 הלכות ממתינות — האם המחלץ נדיב מדי על החלטות-ועדה,
+> או שזה ספציפי לתיק הזה?" התשובה: **שיטתי, לא ספציפי — אבל הנדיבות מוצדקת.**
+>
+> **מתכתב עם [`00-final-synthesis.md`](00-final-synthesis.md):** הנתונים כאן **מחזקים** את הכרעת-הסינתזה
+> ("לא לחתוך") ומוסיפים שני דברים שלא היו לה: (א) כימות חוצה-קורפוס של ה-`application`, (ב) ממצא
+> חדש וגדול יותר — `nli_unsupported` על הפסיקה החיצונית. ראה §4 (יישוב-מתח) ו-§5 (דרכי-פעולה).
+
+---
+
+## 1. הממצא המרכזי — `application` הוא תופעה שיטתית של החלטות-ועדה
+
+הפרדנו את הקורפוס לפי `authority` (נגזר דטרמיניסטית מ-`precedent_level`, [02-data-model §162](../spec/02-data-model.md)):
+`binding` = פסיקת-עליון/מנהלי · `persuasive` = ועדת-ערר מחוזית.
+
+```text
+source                         rows   rt=application   nli_unsupported   any-flag
+court (binding)               3,603        0.2%            39.7%           45.4%
+committee (persuasive)        1,842       13.7%            25.0%           30.5%
+```
+
+- **`rule_type='application'` הוא כמעט-בלעדית של ועדות:** 13.7% מול 0.2%. מתוך 258 רשומות-`application`
+  בכל הקורפוס, **~252 מגיעות מהחלטות-ועדה.** פער של פי-~70.
+- **עקבי בין תחומים** (לא עניין-שמאות נקודתי):
+```text
+committee, rt=application by practice_area:
+  rishuy_uvniya     13.5%      betterment_levy   14.1%      compensation_197   18.8%
+```
+
+**הפרשנות (תואם [02-data-model §163](../spec/02-data-model.md)):** `application` = "החלה תלוית-עובדות —
+לרוב לא-הלכה". ועדת-ערר היא גוף מיישֵם: היא לא יוצרת הלכה (רק בית-משפט עושה זאת), אלא **מיישמת
+פסיקת-עליון קיימת** (לוסטרניק, דלי-דליה) על עובדות-התיק. לכן חילוץ מהחלטת-ועדה מייצר באופן מובנה
+שיעור גבוה של יישומי-דוקטרינה. **זו תכונה של מקור-הנתונים, לא באג של המחלץ.**
+
+## 2. 8508-03-24 — מייצג בקצה-העליון, לא חריג
+
+```text
+דירוג 8508-03-24:   7 מתוך 45 תיקי-ועדה (≥15 רשומות)   ·   27% application (rt או flag)
+חציון תיקי-הוועדה:                                          11.9%
+מעליו: 1001-02-19 (40%) · 1044-08-22 (37%) · 9002-24 (33%) · 1007-01-25 (30%)
+```
+
+8508 הוא ~פי-2.3 מהחציון (רבעון-עליון), אבל מה שבולט בו הוא בעיקר ש**הוא התיק הארוך בקורפוס**
+(111 רשומות) — אז 27% נותן 30 פריטי-`application` במספר מוחלט, הגבוה בקורפוס. כלומר: **כמות גבוהה,
+שיעור גבוה-אך-נורמלי.** אין כאן פתולוגיה ייחודית לתיק.
+
+## 3. מנגנון-הניתוב הקיים (איך זה כבר מטופל ב-UI)
+
+הפיצול בתור-ההלכות (`/precedents` → "תור הלכות") אינו לפי `exclude_low_quality`, אלא לפי
+`isExtractionFixItem(h) = (quality_flags.length>0) && !panel_round`
+([web-ui/.../precedent-library.ts:652](../../web-ui/src/lib/api/precedent-library.ts#L652)):
+
+```text
+8508-03-24, 71 ממתינות (מצב 2026-06-20):
+  bucket                          #     panel?   →"להכרעתך"   →"דורש תיקון-חילוץ"
+  clean (ללא דגל)                40     15/40        40              0
+  application                    23      0/23         0             23
+  nli_unsupported                 4      4/4          4              0
+  thin_restatement                4      0/4          0              4
+```
+
+**משמעות:** פריטי-`application` (חסרי-פאנל) כבר מנותבים ל**"דורש תיקון-חילוץ"** — **מחוץ** לתור-ההכרעה
+של היו"ר. כלומר המערכת כבר מסננת אותם מהעומס-הידני. (הפאנל התלת-מודלי, [halacha_panel_approve.py:191](../../scripts/halacha_panel_approve.py#L191),
+מטפל רק בדליים `clean`+`nli`; `application` ו-`defect` עוקפים אותו במכוון.)
+
+## 4. ⚠️ יישוב-המתח מול הסינתזה — `application` ≠ "רעש"
+
+זו הנקודה הקריטית להעברה. **אסור** לתרגם "13.7% application" ל"13.7% רעש לחיתוך". המבחן בסינתזה
+(§2 שם) כבר הוכיח שחיתוך-אגרסיבי על 8508 השמיד את **לוסטרניק** ו-~22 עקרונות-ליבה. רוב פריטי-ה-`application`
+הם בדיוק יישומי-הדוקטרינה הללו — **בני-ציטוט שהכותב צריך**. דוגמאות אמיתיות מ-8508 שמסומנות `application`:
+
+```text
+"ציפיות הנובעות אך ממיקומם של המקרקעין... אין לנטרלן"        ← לוסטרניק מיושמת — לשמור!
+"בחישוב שווי במצב קודם יש לכלול ציפיות כלליות... ולא ספציפיות" ← ליבת חישוב היטל-ההשבחה
+"קביעת מקדם מצויה בליבת שיקול-דעת השמאי, והוועדה לא תתערב"     ← סטנדרט אי-התערבות מיושם
+```
+
+לכן: **הנתון הזה תומך ב"שמור-בספק" של עמוד-1/2 בסינתזה.** המסקנה הנכונה אינה "לסנן application" אלא
+"`application` מאשר שעקרוני-הוועדה הם persuasive-יישומיים — לדרג אותם נמוך באחזור (רמה B), לא למחוק
+אותם (רמה A)". `importance=0` ל-8508 כבר משקיע אותם ממילא.
+
+## 5. ⭐ ממצא-לוואי גדול יותר — `nli_unsupported` על הפסיקה החיצונית
+
+```text
+nli_unsupported:   court (binding) 39.7%   ≫   committee 25.0%
+any-flag:          court 45.4%   ·   committee 30.5%
+```
+
+**כמעט מחצית מעקרוני-הפסיקה-החיצונית נושאים דגל-איכות, בעיקר `nli_unsupported`** (הכלל אינו נגזר
+לוגית מהציטוט התומך שלו, [halacha_quality.py:282](../../mcp-server/src/legal_mcp/services/halacha_quality.py#L282)).
+זה **מגמד מספרית** את סוגיית-ה-`application`, ונוגע ישירות ל"רמה A = ניקוי-רעש" של הסינתזה:
+**מהו ה"רעש" שמנקים?** הדגל הדומיננטי הוא `nli`, והוא מרוכז בפסיקה, לא בוועדות.
+
+שתי השערות מתחרות, שצריך להכריע ביניהן **לפני** שמשתמשים ב-`nli` כמסנן-רעש ברמה A:
+- **(א) הבודק מחמיר מדי** — סף-ה-NLI חותך יישורים לגיטימיים → 40% הם false-positives, וה"רעש" מדומה.
+- **(ב) החילוץ-מהפסיקה לקוי** — ציטוטים תומכים שלא מיישרים לכלל → בעיית-חילוץ אמיתית בקנה-מידה.
+
+ההכרעה משנה את כל אסטרטגיית רמה-A. **אנו ממליצים לאמת זאת על מדגם-זהב לפני כל שימוש ב-`nli` כסיגנל.**
+
+---
+
+## 6. דרכי-פעולה מוצעות (לסוכן-הקורפוס)
+
+ממוין מהשמרני לאגרסיבי. ההמלצה שלנו: **B כברירת-מחדל + C כעבודה-מקבילה**; להימנע מ-A ומ-D.
+
+| # | פעולה | טיעון | סיכון | המלצה |
+|---|-------|-------|-------|-------|
+| **A** | לכוונן את ה-prompt לדכא `application` מוועדות במקור | מטפל-בשורש (G1); חוסך 14% רשומות | **גבוה** — סותר את מבחן-8508; משמיד יישומי-לוסטרניק בני-ציטוט | ✗ לא |
+| **B** | להשאיר את החילוץ; לסמוך על הניתוב הקיים (`application`→"תיקון", מחוץ לתור-היו"ר) + לדרג נמוך ב-RRF | תואם-סינתזה (שמור-בספק + דרג-בזמן-אחזור); אפס סיכון-אובדן | הרעש נשאר ב-DB (אחסון בלבד) | ✓ **כן — ברירת-מחדל** |
+| **C** | לחקור קודם את `nli_unsupported` (40% פסיקה): מדגם-זהב, להכריע (א) מחמיר-מדי מול (ב) חילוץ-לקוי | זה הסיגנל הגדול; הכרחי לפני שמגדירים "רעש" ברמה A | דורש מדגם מתויג-ידנית | ✓ **כן — במקביל** |
+| **D** | להפסיק חילוץ-עקרונות מוועדות לגמרי | רוב עקרוני-הוועדה הם שכתוב-persuasive של עליון | **קיצוני** — נוגד 07-learning §61 (ועדות ברות-ציטוט במכוון); נוגע INV-LRN | ✗ לא (אלא בהכרעת-יו"ר מפורשת) |
+
+### ההמלצה המזוקקת
+1. **לא לגעת בחילוץ-מוועדות** — הנתון מאשר שהנדיבות מוצדקת; `application`=יישום-בר-ציטוט, לא זבל. עקבי עם
+   הכרעת-הסינתזה "לא לחתוך".
+2. **רמה B עושה את העבודה** — `importance` boost ב-RRF מטביע את עקרוני-הוועדה ה-persuasive מתחת
+   לפסיקה-המחייבת, בלי למחוק דבר. 8508 (`importance≈0`) שוקע ממילא.
+3. **להעביר את ה-`nli` לראש תור-המחקר** — לפני שמשתמשים בו כמסנן-רעש ברמה A, לאמת אם 40% אמיתי.
+
+---
+
+> **מקור-הנתונים:** `GET /api/halachot?limit=100000` (5,489 שורות) + per-case `?case_law_id=…`.
+> ניתן לשחזר את כל המספרים מהשאילתות האלו. הקאנון הוא live — שיעורים ינועו ככל שהדריינר/פאנל רצים.

docs/spec/07-learning.md

@@ -40,7 +40,7 @@
 3. **C — deep-read (נקודתי):** voice-XXXX.md — worked example לתיק-מופת.
 
 ### 0.3 הצינור החוזר per-final (7 שלבים)
-`mark-final` → [1] INTAKE (snapshot של הטיוטה) → [2] PAIRING (בלוק↔בלוק) → [3] ALIGNMENT (diff פר-בלוק) → [4] DISTILLATION (מפריד סגנון↔מהות) → [5] CURATION (Hermes + שער-יו"ר) → [6] FEEDBACK (ניתוב לערוץ A/B/C) → [7] MEASUREMENT (מדד-מרחק-סגנון).
+`mark-final` → [1] INTAKE (snapshot של הטיוטה) → [2] PAIRING (בלוק↔בלוק) → [3] ALIGNMENT (diff פר-בלוק) → [4] DISTILLATION (מפריד סגנון↔מהות) → [5] CURATION (Hermes + שער-יו"ר) → [6] FEEDBACK (ניתוב לערוץ A/B/C) → [7] MEASUREMENT (מדד-מרחק-סגנון + snapshot-held-out פרוספקטיבי, §0.7).
 
 ### 0.4 ניהול ב-UI
 `/methodology` = **עורך-הפרופיל היחיד** (declarative: יחסי-זהב, כללי-דיון, צ׳קליסטים, ביטויי-מעבר, אנטי-דפוסים, voice-invariants). `/training` = **שולחן-הלמידה** (קורפוס, פורטרט-סגנון, השוואת draft↔final, curator, מדד-מרחק, פנקס-התאמה).
@@ -48,7 +48,7 @@
 **שער-אישור אחד · טרנזקציית-כותב אחת (INV-IA3 → [X17](X17-information-architecture.md)):** ל-`decision_lesson` יש **סטטוס-יחיד** שקובע "זורם-לכותב" — `review_status='approved'` (INV-LRN1/G10). הדגל `applied_to_skill` **הוסר** (היה אינפורמטיבי-בלבד, נכתב-לשומקום → בלבל את היו"ר ב"שני שערים"; גל-2 #131). לקח שהיו"ר מחבר ידנית נוצר כבר כ-`approved`; לקח-פאנל נוצר כ-`proposed` וממתין לשער. promote של זוג draft↔final מטמיע את הלקחים/הביטויים שהיו"ר בחר **דרך appeal_type_rules בטרנזקציה אחת נעולה (FOR UPDATE)** — מסלול-כתיבה-יחיד, ללא read-modify-write מתפצל מול עורך-המתודולוגיה (MET-2/3, להלן G2 הפרות-ידועות).
 
 ### 0.5 Invariants חדשים
-**INV-LRN4 (ניגוד-אמת → G10/G9):** למידת-קול מבוססת **pairing draft↔final ברמת-בלוק**, לא קריאת-final בלבד. כל החלטה אינה "סגורה" עד שהושוותה מול הסופי; כל סופי מנותח מול הטיוטה. נשמר פנקס-התאמה (`draft_final_pairs`) עם מצב-חיים `draft_done → final_received → analyzed → lessons_folded`.
+**INV-LRN4 (ניגוד-אמת → G10/G9):** למידת-קול מבוססת **pairing draft↔final ברמת-בלוק**, לא קריאת-final בלבד. כל החלטה אינה "סגורה" עד שהושוותה מול הסופי; כל סופי מנותח מול הטיוטה. נשמר פנקס-התאמה (`draft_final_pairs`) עם מצב-חיים `draft_done → final_received → analyzed → lessons_folded`. ההכללה (האם הלמידה משפרת תיקים שלא-נלמדו) נמדדת פרוספקטיבית — §0.7.
 *מקורות:* imitation-learning-from-expert-edits · contrastive personalization (arxiv 2504.08745) · author-profiling. *סטטוס: verified.*
 
 **INV-LRN5 (טוהר-הקול → G4/G11):** שכבת-ידע-הקול (voice-fingerprint, style_patterns, exemplars) **לא תכיל הלכות/עובדות ספציפיות** — רק סגנון ושיטה. מהות מנותבת ל-precedent_library/halacha. ה-distillation מפריד במקור.
@@ -62,9 +62,18 @@
 3. **בדיקת-ציטוטים** — `extract_internal_citations` מקשר את הפסיקה שההחלטה מצטטת לספרייה; כל ציטוט שאינו בספרייה **מסומן אוטומטית** כ-`missing_precedent` (open) להעלאה ע"י היו"ר.
 4. הציטוטים-המקושרים מזינים את **לולאת-ה-corroboration** (X11): ציטוט-נכנס מההחלטה שלנו מחזק את ההלכות של התקדים המצוטט (`corroboration_rebuild`).
 ואז שני שלבים אוטומטיים נפרדים (`run-learning` / `run-halacha`) המעירים worker מקומי (claude/DeepSeek/Gemini מקומיים בלבד):
-- **למידה:** `ingest_final_version` (Opus distillation) → **פאנל-סגנון דו-סוכני** (DeepSeek+Gemini, "למידה כפולה") שמצביע על כל לקח-style_method; הסכמה 2/2 → `decision_lesson` (`source=panel:deepseek+gemini`); פיצול → ליו"ר.
+- **למידה:** הצינור (`final_learning_pipeline.py`) רץ בסדר `enroll_style_corpus` (יצירת רשומת-הקורפוס, מהיר — **ראשון** מאז #159 כדי שהקורפוס יהיה זמין לפני הדיסטילציה הארוכה) → `ingest_final_version` (Opus distillation) → **פאנל-סגנון דו-סוכני** (DeepSeek+Gemini, "למידה כפולה") שמצביע על כל לקח-style_method; הסכמה 2/2 → `decision_lesson` (`source=panel:deepseek+gemini`) **שזורם אוטומטית לכותב** כ-`approved` (שער-מדורג, INV-LRN1) — הפיך (veto-יו"ר ב-/training); פיצול → ליו"ר. **במקביל** (יקיצת `final_learning_*`) האוצֵר ממשיך ל-§A ורושם ממצאי-`source='curator'` (ערוץ ג׳, §1.1).
 - **הלכות:** `extract_internal_citations` → `precedent_extract_halachot` → `corroboration_rebuild` → **פאנל-הלכות תלת-סוכני** (`halacha_panel_approve.py --apply`).
-שני הפאנלים **הפיכים** (גיבוי-CSV ל-`data/audit/`) ומסלימים מחלוקות. ההטמעה הסופית ל-`SKILL.md`/`legal-decision-lessons.md` נשארת **אישור-יו"ר ידני** (INV-LRN1/G10) — הפאנל יוצר *הצעות* בלבד.
+שני הפאנלים **הפיכים** (גיבוי-CSV ל-`data/audit/`) ומסלימים מחלוקות. ההטמעה ל-`SKILL.md`/`legal-decision-lessons.md` ולכל **מהות** נשארת **אישור-יו"ר ידני קשיח** (INV-LRN1/G10); לקחי-**סגנון** בקונצנזוס זורמים אוטומטית-והפיך לכותב.
+
+### 0.7 מדד-ההכללה הפרוספקטיבי (held-out trend — "מסלול A")
+**השאלה שזה עונה:** האם הלמידה באמת *מכלילה* — כלומר משפרת טיוטות של תיקים **שלא נלמדו** — ולא רק "משננת" תיקים שכבר ראינו. מדד-מרחק-הסגנון (שלב [7]) על תיק שלקחיו כבר הוטמעו אינו held-out; מדידה רטרואקטיבית בלתי-אפשרית כי הלקחים נשמרים `appeal_type_rules` (`universal`) **ללא תיוג-מקור** → אין leave-one-out. לכן המדידה **פרוספקטיבית**: נלכדת ברגע היחיד שבו היא נקייה.
+
+**המנגנון (אוטומטי, ב-`final/upload`, ללא LLM):** מיד אחרי פתיחת הזוג `draft_final_pair` אבל **לפני** הטמעת-לקחי-התיק (ה-fold הוא שלב-`promote` ידני נפרד ב-`/training`, §0.4), נלכד snapshot של `style_distance` (`anti_pattern_total`, סטיית-יחסי-זהב, `change_percent`) יחד עם **גודל-בריכת-הלקחים** באותו רגע (מספר `discussion_rules`+`transition_phrases` ב-`universal`). מכיוון שהטיוטה נכתבה עם הבריכה ה**קודמת** בלבד, כל שורה היא נקודת-נתון held-out נקייה: *"עם N לקחים מצטברים, הטיוטה שלנו על תיק שלא-נראה קיבלה ציון X"*.
+
+**הקריאה:** טבלת `style_distance_history` (append-only) + `GET /api/learning/style-distance-history`. ירידה ב-`anti_pattern_total`/`change_percent` ככל ש-N גדל = **הוכחה מתגלגלת שהלמידה מכלילה** (INV-LRN4 — זהו משטח-המגמה של "ניגוד-האמת"). **אזהרת-פרשנות:** `change_percent` מערבב סגנון עם שלמות-תוכן (לפעמים היו"ר מכפילה אורך כי חסרה מהות, לפעמים חותכת) → `anti_pattern_total` הוא הסיגנל הנקי-יותר לסגנון. שימוש-חוזר בשירות `style_distance` ובבריכת `appeal_type_rules` — אין מסלול-מדד מקביל (G2).
+
+> **חלון נקי חד-פעמי:** תיק שכבר `lessons_folded` פספס את חלון ה-held-out שלו — אין backfill. הטבלה מתמלאת קדימה מהסופי הבא.
 
 ---
 
@@ -87,10 +96,24 @@
   (`hermes-curator.md:60-70`).
 - מזהה **3–5 דפוסים/פערים** חדשים, כל ממצא מתויג `[סגנון]` / `[מבנה]` /
   `[לקסיקון משפטי]` / `[טבלאי]` (`hermes-curator.md:99-108`).
-- **מציע** — comment ב-Paperclip + רישום כל ממצא כ-`decision_lesson` דרך
+- **מציע** — comment + interaction בערוץ-הפלטפורמה, וגם **רושם כל ממצא מבנית** כ-`decision_lesson`
-  `POST /api/training/corpus/{corpus_id}/lessons` (`source:"curator"`) שמופיע ב-UI
+  (`source='curator'`, `review_status='proposed'`) דרך כלי-ה-MCP `record_curator_findings`
-  תחת הטאב "מה למדנו" (`hermes-curator.md:73-96`).
+  (`hermes-curator.md` §A.5b → `tools/workflow.py::record_curator_findings` → `db.add_decision_lesson`).
-- **אינו מעדכן** קבצים בעצמו (skills/, lessons.py, DB) — רק מציע (`hermes-curator.md:125-130`).
+  הממצאים מופיעים ב-`/training` טאב "אוצֵר" (ערוץ ג׳) ועוברים את שער-היו"ר (INV-LRN1).
+  > **דריפט-מימוש שנסגר (2026-06-28):** עד גל זה הסוכן כתב comment בלבד והספ תיאר endpoint שלא נקרא —
+  > הממצאים האיכותיים אבדו, וכרטיס-האוצֵר (`get_curator_stats`) ספר `source='curator'` שאיש לא כתב → 0.
+  > `record_curator_findings` מממש את INV-LRN3 בפועל.
+- **אינו מאשר ואינו מטמיע** — read-only על התוכן; רישום-הממצא הוא *הצעה* מגודרת-שער, לא שינוי-קול.
+  אינו עורך `skills/` / `lessons.py` / שכבת-הקול בעצמו (`hermes-curator.md` §"כללים כלליים").
+- **שני ערוצי-למידה נפרדים על אותו סופי:** ערוץ-האוצֵר (כאן, qualitative, `source='curator'`) ו**פאנל-הסגנון**
+  הדו-סוכני שהצינור מפעיל (§0.6 — DeepSeek+Gemini, `source='panel:deepseek+gemini'`). שניהם
+  `decision_lessons` ממתיני-שער; ערוץ הדיסטילציה (`appeal_type_rules`, §0.4) הוא השלישי.
+
+> **נסגר (TaskMaster #159, 2026-06-28):** ה-`PIPELINE-WAKE BRANCH` (`hermes-curator.md`) מבדיל כעת
+> `final_learning_*` מ-`final_halacha_*`: על **learning** הסוכן מפעיל את הצינור ברקע ו**ממשיך ל-§A**
+> (מצב AUTO — רושם ממצאי-`curator`, מדלג על interaction), על **halacha** יוצא מיד כקודם. המירוץ מול
+> `enroll_style_corpus` נסגר בכך ש-enroll רץ **ראשון** בצינור (§0.6) — הקורפוס קיים תוך שניות, הרבה לפני
+> שהסוכן/ה-§A מסיים את ניתוח-ה-LLM; `record_curator_findings` כולל גם retry כרשת-ביטחון.
 
 ### 1.2 לולאת-פידבק-היו"ר (capture → ניתוח שבועי → לקחים)
 
@@ -162,13 +185,12 @@
 
 ## 3. Invariants של התחום
 
-### INV-LRN1: עדכון-ידע דורש אישור-יו"ר ידני — אין auto-commit (governance →G10)
+### INV-LRN1: עדכון-ידע דורש שער-יו"ר — **שער מדורג** לפי סיכון (governance →G10)
-**כלל:** מנגנוני-הלמידה (Hermes, ניתוח-פידבק שבועי) **מציעים בלבד**. כל שינוי ב-
+**כלל (מעודכן 2026-06-28, הכרעת-יו"ר):** השער **מדורג לפי סיכון-התוכן**, לא אחיד:
-[SKILL.md](../../skills/decision/SKILL.md) או ב-[legal-decision-lessons.md](../legal-decision-lessons.md)
+- **מהות** (הלכה / תקדים / עובדה / כל שינוי ב-[SKILL.md](../../skills/decision/SKILL.md) או ב-[legal-decision-lessons.md](../legal-decision-lessons.md)) → **שער-קשיח**: בחינה ואישור ידניים של היו"ר/חיים ואז commit ידני — **לעולם לא auto-committed**. פאנל-ההלכות התלת-סוכני מציע בלבד.
-מחייב **בחינה ואישור ידניים של היו"ר/חיים** ואז commit ידני — **לעולם לא auto-committed**.
+- **סגנון** (`decision_lessons` בקטגוריות style/structure/lexicon/tabular, וכן discussion_rules / transition_phrases / anti_patterns) שעבר **קונצנזוס-פאנל 2/2** → **זורם אוטומטית לכותב** (`review_status='approved'`), **הפיך**: היו"ר רואה ויכול **לבטל בדיעבד** ב-/training. זהו שינוי-קול נמוך-סיכון (טהור-מהות לפי INV-LRN5), בקרת-איכות מובנית (2 מודלים), והפיכוּת — ולכן עדיין "תחת בקרת-המשתמש" במובן NCSC/CEPEJ (veto, לא אישור-מראש על כל פריט).
-Hermes כותב comment + `decision_lesson`, לא קבצים; ה-CEO השבועי כותב לקובץ אך הצעותיו
+
-מאומתות ידנית לפני קיבוע. זהו פֶּאֶט של [INV-G10](00-constitution.md#inv-g10-המערכת-מסייעת--שערים-אנושיים-הם-invariant)
+זהו פֶּאֶט של [INV-G10](00-constitution.md#inv-g10-המערכת-מסייעת--שערים-אנושיים-הם-invariant) על שכבת-הידע: הלמידה כפופה לשיקול-הדעת האנושי — קשיח למהות, הפיך-veto לסגנון. **מימוש:** `scripts/style_lesson_panel.py:_review_status_for` קובע `approved` לקטגוריות-סגנון בקונצנזוס, `proposed` אחרת; הכותב צורך רק `approved` ([db.get_recent_decision_lessons]).
-על שכבת-הידע: גם הלמידה כפופה לשיקול-הדעת האנושי.
 **מקורות:** NCSC/JTC — *Principles & Practices for AI Use in Courts* (human-in-the-loop;
 never replace human judgment) · Council of Europe / CEPEJ (2018, under user control) ·
 Federal Judicial Center — *Judicial Writing Manual* (2d ed.) | סטטוס: verified

docs/spec/X8-field-provenance.md

@@ -35,14 +35,14 @@
 | `case_number` (citation) | CHAIR (חובה) | מפתח idempotency |
 | `full_text`, `extraction_status`, `source_kind` | DETERMINISTIC | — |
 | `case_name`, `court`, `date`, `headnote`, `summary`, `key_quote`, `subject_tags`, `appeal_subtype`, `precedent_level`, `source_type`, `citation_formatted` | CHAIR או OPUS | Opus ממלא רק אם ריק |
-| `is_binding` | CHAIR (default true) | קובע prompt-הלכה |
+| `is_binding` | CHAIR (default true) **אך** DETERMINISTIC-false ל-`source_type=appeals_committee` | קובע prompt-הלכה. סמכות מבנית ([INV-DM7](02-data-model.md#inv-dm7-סיווג-הלכה--סמכות-נגזרת--תפקיד-כלל-מסווג-שני-צירים-לא-enum-אחד)): מקור-ועדה הוא תמיד persuasive — נכפה ל-false ב-`create_external_case_law` ונאכף ע"י `case_law_committee_not_binding_check` |
 | chunks (`content`/`section_type`/`page_number`) | DETERMINISTIC | — |
 | `embedding` (chunks) | Voyage (לא-LLM-reasoning) | ⚠ לא-GENERATED ([gap-audit GAP-09](gap-audit.md)) |
 | כל `halachot` | OPUS | נכנס pending_review |
 
 ### 2ב. החלטה פנימית (`case_law`, source_kind=`internal_committee`)
 כמו 2א, ובנוסף: `case_number` **חובה**; `chair_name`/`district`/`proceeding_type` — CHAIR או OPUS או DERIVED;
-`source_type` = `appeals_committee` (DETERMINISTIC קבוע). placeholder `"(טרם חולץ)"` מסומן ל-chair_name/district
+`source_type` = `appeals_committee` (DETERMINISTIC קבוע); `is_binding` = `false` (DETERMINISTIC קבוע — נכפה ב-`create_internal_committee_decision`, [INV-DM7](02-data-model.md#inv-dm7-סיווג-הלכה--סמכות-נגזרת--תפקיד-כלל-מסווג-שני-צירים-לא-enum-אחד): החלטת-ועדה אינה מחייבת ועדה אחרת ולא את עצמה — תמיד persuasive). placeholder `"(טרם חולץ)"` מסומן ל-chair_name/district
 ריקים ומטופל כריק ע"י ה-extractor.
 
 ### 2ג. מסמך-תיק (`documents`)

docs/superpowers/plans/2026-05-16-hooks-execution.md

@@ -0,0 +1,348 @@
+# Hooks: Case Status Webhooks Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** כאשר status של תיק משתנה ב-legal-ai (e.g. `qa_failed`, `exported`), ה-plugin מקבל webhook, מעדכן את ה-issue ב-Paperclip, ומעיר את ה-CEO במקרה הצורך.
+
+**Architecture:** Legal-ai REST API קורא `pc_request("POST", "/api/plugins/marcusgroup.legal-ai/webhooks/case-status", ...)` אחרי כל שינוי status. ה-plugin מטפל ב-`onWebhook()` ומגיב: מוסיף תגובה לissue, מעיר CEO אם QA נכשל.
+
+**Tech Stack:** TypeScript (plugin-legal-ai), Python/FastAPI (legal-ai web), `@paperclipai/plugin-sdk@2026.325.0`, `httpx` (Python), `pc_request` helper.
+
+---
+
+## File Map
+
+| Action | File |
+|--------|------|
+| Modify | `plugin-legal-ai/src/worker.ts` — add `onWebhook()` to `definePlugin({})` |
+| Modify | `plugin-legal-ai/plugin.json` — add `"webhooks.receive"` capability |
+| Modify | `plugin-legal-ai/src/manifest.ts` — add webhook capability |
+| Modify | `legal-ai/web/app.py` — emit webhook after `PUT /api/cases/{case_number}` |
+| Modify | `legal-ai/web/paperclip_api.py` — add `emit_webhook()` helper |
+
+---
+
+## Task 1: Add `emit_webhook` helper ב-Python
+
+**Files:**
+- Modify: `legal-ai/web/paperclip_api.py`
+
+- [ ] **Step 1: קרא את הקובץ הקיים**
+
+```bash
+head -90 /home/chaim/legal-ai/web/paperclip_api.py
+```
+
+- [ ] **Step 2: הוסף את ה-helper בסוף הקובץ**
+
+פתח `/home/chaim/legal-ai/web/paperclip_api.py` והוסף אחרי הפונקציה `pc_request`:
+
+```python
+async def emit_case_status_webhook(
+    case_number: str,
+    old_status: str,
+    new_status: str,
+    company_id: str | None = None,
+    run_id: str | None = None,
+) -> None:
+    """Notify the Paperclip plugin that a case status changed.
+
+    Fire-and-forget: logs errors but never raises, so callers aren't blocked.
+    """
+    try:
+        await pc_request(
+            "POST",
+            "/api/plugins/marcusgroup.legal-ai/webhooks/case-status",
+            json={
+                "caseNumber": case_number,
+                "oldStatus": old_status,
+                "newStatus": new_status,
+                "companyId": company_id,
+                "timestamp": datetime.utcnow().isoformat() + "Z",
+            },
+            run_id=run_id,
+            timeout=5.0,
+        )
+    except Exception as exc:
+        logger.warning("emit_case_status_webhook failed: %s", exc)
+```
+
+> **הערה:** `datetime` ו-`logger` כבר מיובאים ב-`app.py`. בדוק שהם מיובאים גם ב-`paperclip_api.py` — אם לא, הוסף `from datetime import datetime` ו-`import logging; logger = logging.getLogger(__name__)` בראש הקובץ.
+
+- [ ] **Step 3: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add web/paperclip_api.py
+git commit -m "feat: add emit_case_status_webhook helper"
+```
+
+---
+
+## Task 2: צרף webhook לendpoint `PUT /api/cases/{case_number}`
+
+**Files:**
+- Modify: `legal-ai/web/app.py`
+
+- [ ] **Step 1: מצא את ה-endpoint**
+
+```bash
+grep -n "PUT\|case_number\|update_case" /home/chaim/legal-ai/web/app.py | head -30
+```
+
+- [ ] **Step 2: קרא את הenable endpoint המלא**
+
+זהה את הסקציה המלאה של ה-endpoint ואת הייבוא הקיים.
+
+- [ ] **Step 3: הוסף import ל-emit_webhook**
+
+בראש `app.py`, בסקציית ה-imports מ-`paperclip_api`:
+```python
+from .paperclip_api import pc_request, emit_case_status_webhook
+```
+
+- [ ] **Step 4: הוסף webhook emit בתוך הendpoint**
+
+אחרי שהקוד מעדכן את ה-case (לפני ה-`return`), הוסף:
+
+```python
+# Notify plugin about status change (fire-and-forget)
+if updates.get("status") and old_status != updates["status"]:
+    background_tasks.add_task(
+        emit_case_status_webhook,
+        case_number=case_number,
+        old_status=old_status,
+        new_status=updates["status"],
+        company_id=str(case.get("company_id")),
+    )
+```
+
+> אם ה-endpoint כבר מקבל `background_tasks: BackgroundTasks` — השתמש בו. אם לא, הוסף `background_tasks: BackgroundTasks` לחתימת הפונקציה. הוסף `from fastapi import BackgroundTasks` ל-imports.
+
+- [ ] **Step 5: שמור את ה-`old_status` לפני ה-update**
+
+בתחילת ה-endpoint handler, לפני קריאת ה-DB update:
+```python
+old_status = (await db.get_case(case_number) or {}).get("status", "")
+```
+
+- [ ] **Step 6: בדיקה בסיסית — שלח PUT ידני**
+
+```bash
+curl -s -X PUT https://legal-ai.nautilus.marcusgroup.org/api/cases/1130-25 \
+  -H "Content-Type: application/json" \
+  -d '{"status": "in_progress"}' | jq .status
+```
+
+בדוק ב-Paperclip logs שהwebhook נשלח (עדיין לא מטופל בצד ה-plugin):
+```bash
+pm2 logs paperclip --lines 20
+```
+
+- [ ] **Step 7: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add web/app.py
+git commit -m "feat: emit case-status webhook on PUT /api/cases/:case"
+```
+
+---
+
+## Task 3: הוסף `onWebhook()` ל-plugin
+
+**Files:**
+- Modify: `plugin-legal-ai/src/worker.ts`
+
+- [ ] **Step 1: קרא את `definePlugin({})` הקיים**
+
+```bash
+grep -n "definePlugin\|onWebhook\|onHealth\|onShutdown" /home/chaim/plugin-legal-ai/src/worker.ts
+```
+
+- [ ] **Step 2: הוסף את `onWebhook` handler**
+
+בתוך הobject שמועבר ל-`definePlugin({})`, אחרי `setup(ctx)`:
+
+```typescript
+onWebhook: async (input) => {
+  const { endpointKey, payload, companyId } = input as {
+    endpointKey: string;
+    payload: {
+      caseNumber: string;
+      oldStatus: string;
+      newStatus: string;
+      companyId: string;
+      timestamp: string;
+    };
+    companyId: string;
+  };
+
+  if (endpointKey !== "case-status") return;
+
+  const { caseNumber, oldStatus, newStatus } = payload;
+  ctx.logger.info(`Webhook: ${caseNumber} ${oldStatus} → ${newStatus}`);
+
+  // Find the Paperclip issue linked to this case
+  const stateKey = `case:${caseNumber}`;
+  const issueId = await ctx.state.get({ companyId }, stateKey);
+  if (!issueId) {
+    ctx.logger.warn(`No issue found for case ${caseNumber}`);
+    return;
+  }
+
+  const statusLabels: Record<string, string> = {
+    in_progress: "🔄 בעבודה",
+    drafted: "✍️ טיוטה מוכנה",
+    qa_failed: "❌ QA נכשל",
+    exported: "📄 יוצא ל-DOCX",
+    reviewed: "✅ נבדק",
+    final: "🎯 סופי",
+  };
+
+  const label = statusLabels[newStatus] ?? newStatus;
+
+  // Post a status comment on the issue
+  await ctx.issues.createComment({
+    issueId: issueId as string,
+    body: `**עדכון סטטוס:** ${label} (היה: ${oldStatus})`,
+  });
+
+  // Wake CEO if QA failed
+  if (newStatus === "qa_failed") {
+    const companies = await ctx.companies.list();
+    const company = companies.find((c) => c.id === companyId);
+    if (!company) return;
+
+    const CEO_IDS: Record<string, string> = {
+      "42a7acd0-30c5-4cbd-ac97-7424f65df294": "752cebdd-6748-4a04-aacd-c7ab0294ef33",
+      "8639e837-4c9d-47fa-a76b-95788d651896": "cdbfa8bc-3d61-41a4-a2e7-677ec7d34562",
+    };
+    const ceoId = CEO_IDS[companyId];
+    if (ceoId) {
+      await ctx.agents.invoke(ceoId, companyId, {
+        prompt: `תיק ${caseNumber} נכשל בבדיקת QA. עיין בתוצאות QA ותקן את הבעיות.`,
+        reason: "qa_failed webhook",
+      });
+    }
+  }
+},
+```
+
+> **הערה:** `ctx` חייב להיות נגיש ב-`onWebhook`. אם `ctx` מוגדר בתוך `setup()` בלבד — הוצא אותו ל-closure חיצוני של ה-plugin object (ראה את הדפוס הקיים ב-`worker.ts`).
+
+- [ ] **Step 3: בדק TypeScript**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npx tsc --noEmit
+```
+
+Expected: 0 errors.
+
+- [ ] **Step 4: Build**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npm run build
+```
+
+Expected: `dist/worker.js` נוצר ללא שגיאות.
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd /home/chaim/plugin-legal-ai
+git add src/worker.ts
+git commit -m "feat: add onWebhook handler for case-status events"
+```
+
+---
+
+## Task 4: הוסף capability ל-`plugin.json` ול-`manifest.ts`
+
+**Files:**
+- Modify: `plugin-legal-ai/plugin.json`
+- Modify: `plugin-legal-ai/src/manifest.ts`
+
+- [ ] **Step 1: הוסף `"webhooks.receive"` ל-capabilities**
+
+ב-`plugin-legal-ai/plugin.json`, בarray `"capabilities"`, הוסף:
+```json
+"webhooks.receive"
+```
+
+- [ ] **Step 2: הוסף גם ב-`manifest.ts`**
+
+```bash
+grep -n "capabilities\|webhooks" /home/chaim/plugin-legal-ai/src/manifest.ts
+```
+
+הוסף `"webhooks.receive"` לarray שם.
+
+- [ ] **Step 3: Re-install plugin**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npm run build
+npx paperclipai plugin uninstall marcusgroup.legal-ai \
+  --api-base http://localhost:3100 --api-key pcapi_legal_install_key_2026
+npx paperclipai plugin install /home/chaim/plugin-legal-ai \
+  --api-base http://localhost:3100 --api-key pcapi_legal_install_key_2026
+pm2 restart paperclip
+```
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /home/chaim/plugin-legal-ai
+git add plugin.json src/manifest.ts
+git commit -m "feat: add webhooks.receive capability to plugin manifest"
+```
+
+---
+
+## Task 5: בדיקה end-to-end
+
+- [ ] **Step 1: Deploy legal-ai**
+
+```bash
+cd /home/chaim/legal-ai
+git push origin main
+# המתן לבנייה (~2-4 דקות)
+```
+
+- [ ] **Step 2: שנה סטטוס תיק**
+
+```bash
+curl -s -X PUT https://legal-ai.nautilus.marcusgroup.org/api/cases/1130-25 \
+  -H "Content-Type: application/json" \
+  -d '{"status": "qa_failed"}' | jq .status
+```
+
+- [ ] **Step 3: בדוק שהתגובה הוספה ל-Paperclip issue**
+
+```bash
+# מצא את ה-issue הקשור לתיק 1130-25
+pm2 logs paperclip --lines 30 | grep "1130-25\|webhook\|qa_failed"
+```
+
+Expected: תגובה "❌ QA נכשל" הוספה לissue. CEO הועיר.
+
+- [ ] **Step 4: בדוק שינוי סטטוס שגרתי (לא QA)**
+
+```bash
+curl -s -X PUT https://legal-ai.nautilus.marcusgroup.org/api/cases/1130-25 \
+  -H "Content-Type: application/json" \
+  -d '{"status": "drafted"}' | jq .status
+```
+
+Expected: תגובה "✍️ טיוטה מוכנה" בissue. CEO **לא** הועיר.
+
+---
+
+## אימות סופי
+
+| בדיקה | פקודה | תוצאה מצופה |
+|-------|-------|-------------|
+| QA נכשל → CEO מועיר | `PUT status=qa_failed` | תגובה + agent invocation |
+| exported → תגובה בלבד | `PUT status=exported` | תגובה בלבד |
+| שינוי ללא status | `PUT title=...` | שום webhook |
+| תיק ללא issue | webhook לתיק חדש | לוג warning, ללא crash |

docs/superpowers/plans/2026-05-16-per-agent-claude-md.md

@@ -0,0 +1,306 @@
+# Per-Agent CLAUDE.md Versioning & Validation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** הוסף validation ו-version tracking לקבצי הוראות הסוכנים. כעת, לפני שה-sync script מחיל שינויים, הוא מאמת שכל `instructionsFilePath` קיים. בנוסף, metadata של הסוכן ב-DB יכיל `claude_md_mtime` — השינוי האחרון בקובץ — כדי לזהות drift.
+
+**Architecture:** `sync_agents_across_companies.py` מקבל `--check-instructions` flag שסורק את כל הסוכנים ומדווח על קבצים חסרים/ישנים. ב-`--apply`, מתווספת בדיקת pre-flight שמבטלת את הsync אם קובץ חסר. `agents.metadata` מקבל `claude_md_mtime` עם ה-mtime בפועל של הקובץ.
+
+**Tech Stack:** Python 3.10+, asyncpg, httpx, `os.path.getmtime()`, Paperclip REST API (`PATCH /api/agents/{id}`).
+
+---
+
+## File Map
+
+| Action | File |
+|--------|------|
+| Modify | `legal-ai/scripts/sync_agents_across_companies.py` — `--check-instructions` flag, pre-flight, metadata update |
+
+זה הקובץ היחיד שצריך לגעת בו. כל שאר הלוגיקה קיימת כבר.
+
+---
+
+## Task 1: הוסף `--check-instructions` flag
+
+**Files:**
+- Modify: `legal-ai/scripts/sync_agents_across_companies.py`
+
+- [ ] **Step 1: קרא את החלק של `argparse` בסקריפט**
+
+```bash
+grep -n "argparse\|add_argument\|--verify\|--dry-run\|--apply" \
+  /home/chaim/legal-ai/scripts/sync_agents_across_companies.py | head -20
+```
+
+- [ ] **Step 2: הוסף את הargument**
+
+מצא את הסקציה שמגדירה `args` והוסף:
+
+```python
+parser.add_argument(
+    "--check-instructions",
+    action="store_true",
+    help="Scan all agents' instructionsFilePath and report missing/outdated files",
+)
+```
+
+- [ ] **Step 3: הוסף את הפונקציה `check_instructions()`**
+
+הוסף לפני `async def main()`:
+
+```python
+async def check_instructions(agents: list[dict]) -> bool:
+    """Print a report of all agents' instruction files. Returns True if all OK."""
+    all_ok = True
+    print(f"\n{'Agent':<30} {'File':<60} {'Status':<15} {'Size':>8} {'Modified'}")
+    print("-" * 120)
+
+    for agent in agents:
+        adapter_cfg = agent.get("adapter_config") or {}
+        if isinstance(adapter_cfg, str):
+            import json as _json
+            adapter_cfg = _json.loads(adapter_cfg)
+
+        file_path = adapter_cfg.get("instructionsFilePath", "")
+        name = agent.get("name", agent.get("id", "?"))[:29]
+
+        if not file_path:
+            print(f"{name:<30} {'(none)':<60} {'⚠️  NOT SET':<15}")
+            continue
+
+        if not os.path.exists(file_path):
+            print(f"{name:<30} {file_path[-59:]:<60} {'❌ MISSING':<15}")
+            all_ok = False
+            continue
+
+        stat = os.stat(file_path)
+        size_kb = stat.st_size // 1024
+        mtime = datetime.fromtimestamp(stat.st_mtime).strftime("%Y-%m-%d %H:%M")
+
+        # Compare with DB metadata
+        metadata = agent.get("metadata") or {}
+        if isinstance(metadata, str):
+            import json as _json
+            metadata = _json.loads(metadata)
+
+        db_mtime = metadata.get("claude_md_mtime", "")
+        actual_mtime = str(int(stat.st_mtime))
+        drift = " ⚠️ DRIFT" if db_mtime and db_mtime != actual_mtime else ""
+
+        print(f"{name:<30} {file_path[-59:]:<60} {'✅ OK':<15} {size_kb:>6}KB  {mtime}{drift}")
+
+    print()
+    return all_ok
+```
+
+> `from datetime import datetime` ו-`import os` — בדוק שמיובאים בראש הסקריפט. אם לא, הוסף.
+
+- [ ] **Step 4: הוסף קריאה ל-`check_instructions()` ב-`main()`**
+
+בתוך `async def main()`, אחרי שloading הagents מה-DB:
+
+```python
+if args.check_instructions:
+    all_ok = await check_instructions(master_agents + mirror_agents)
+    sys.exit(0 if all_ok else 1)
+```
+
+- [ ] **Step 5: בדיקה**
+
+```bash
+cd /home/chaim/legal-ai
+python scripts/sync_agents_across_companies.py --check-instructions
+```
+
+Expected: טבלה עם כל הסוכנים, paths, סטטוס ✅/❌.
+
+- [ ] **Step 6: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add scripts/sync_agents_across_companies.py
+git commit -m "feat: add --check-instructions flag to sync script"
+```
+
+---
+
+## Task 2: הוסף pre-flight validation לפני `--apply`
+
+**Files:**
+- Modify: `legal-ai/scripts/sync_agents_across_companies.py`
+
+- [ ] **Step 1: מצא את נקודת הכניסה של `--apply`**
+
+```bash
+grep -n "args.apply\|if.*apply\|apply.*mode" \
+  /home/chaim/legal-ai/scripts/sync_agents_across_companies.py | head -10
+```
+
+- [ ] **Step 2: הוסף pre-flight לפני apply**
+
+בתחילת בלוק `--apply`, לפני כל שינוי:
+
+```python
+if args.apply:
+    # Pre-flight: abort if any agent is missing its instructions file
+    print("🔍 Pre-flight: checking instruction files...")
+    all_ok = await check_instructions(master_agents + mirror_agents)
+    if not all_ok:
+        print("❌ Abort: one or more instruction files are missing. Fix paths before --apply.")
+        sys.exit(1)
+    print("✅ Pre-flight passed.\n")
+    # ... rest of apply logic ...
+```
+
+- [ ] **Step 3: בדיקה — הפעל עם קובץ חסר (סימולציה)**
+
+```bash
+# שנה זמנית path לקובץ שלא קיים
+cd /home/chaim/legal-ai
+python scripts/sync_agents_across_companies.py --dry-run 2>&1 | head -5
+```
+
+Expected: dry-run עובר. אם תנסה `--apply` עם agent שhis file חסר — הsync יבוטל.
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add scripts/sync_agents_across_companies.py
+git commit -m "feat: add pre-flight instruction file validation before --apply"
+```
+
+---
+
+## Task 3: עדכן `agents.metadata` עם `claude_md_mtime`
+
+**Files:**
+- Modify: `legal-ai/scripts/sync_agents_across_companies.py`
+
+- [ ] **Step 1: מצא את `compute_diff()` או `apply_diff()`**
+
+```bash
+grep -n "def compute_diff\|def apply_diff\|def build_patch\|metadata" \
+  /home/chaim/legal-ai/scripts/sync_agents_across_companies.py | head -20
+```
+
+- [ ] **Step 2: הוסף פונקציה `get_claude_md_mtime()`**
+
+```python
+def get_claude_md_mtime(adapter_config: dict) -> str | None:
+    """Return the Unix mtime of the agent's instructionsFilePath, or None if missing."""
+    path = adapter_config.get("instructionsFilePath", "")
+    if not path or not os.path.exists(path):
+        return None
+    return str(int(os.path.getmtime(path)))
+```
+
+- [ ] **Step 3: שלב mtime בbuild של metadata patch**
+
+מצא את המקום שמכין את ה-`metadata` לsync. הוסף:
+
+```python
+# Build metadata patch with claude_md_mtime
+current_metadata = master_agent.get("metadata") or {}
+if isinstance(current_metadata, str):
+    import json as _json
+    current_metadata = _json.loads(current_metadata)
+
+adapter_cfg = master_agent.get("adapter_config") or {}
+if isinstance(adapter_cfg, str):
+    import json as _json
+    adapter_cfg = _json.loads(adapter_cfg)
+
+mtime = get_claude_md_mtime(adapter_cfg)
+if mtime:
+    current_metadata["claude_md_mtime"] = mtime
+    current_metadata["claude_md_last_synced"] = datetime.utcnow().isoformat() + "Z"
+```
+
+כלול את ה-`current_metadata` המעודכן ב-PATCH לAPI.
+
+- [ ] **Step 4: בדוק שה-metadata מתעדכן**
+
+הפעל `--dry-run` וחפש `claude_md_mtime` בoutput:
+```bash
+python scripts/sync_agents_across_companies.py --dry-run 2>&1 | grep -i "mtime\|metadata" | head -10
+```
+
+לאחר `--apply`, בדוק ב-DB:
+```bash
+psql -h localhost -p 54329 -U paperclip -c \
+  "SELECT name, metadata->>'claude_md_mtime' AS mtime FROM agents WHERE metadata->>'claude_md_mtime' IS NOT NULL LIMIT 5" \
+  paperclip
+```
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add scripts/sync_agents_across_companies.py
+git commit -m "feat: track claude_md_mtime in agents.metadata during sync"
+```
+
+---
+
+## Task 4: הוסף `make check-agents` shortcut
+
+**Files:**
+- Modify: `legal-ai/Makefile` (אם קיים) אחרת הוסף alias
+
+- [ ] **Step 1: בדוק אם Makefile קיים**
+
+```bash
+ls /home/chaim/legal-ai/Makefile 2>/dev/null && echo "EXISTS" || echo "MISSING"
+```
+
+- [ ] **Step 2א: אם Makefile קיים — הוסף target**
+
+```makefile
+check-agents:
+	python scripts/sync_agents_across_companies.py --check-instructions
+
+sync-agents-dry:
+	python scripts/sync_agents_across_companies.py --dry-run
+
+sync-agents:
+	python scripts/sync_agents_across_companies.py --apply
+```
+
+- [ ] **Step 2ב: אם Makefile לא קיים — הוסף alias ל-`~/.bashrc`**
+
+```bash
+echo "alias check-agents='cd /home/chaim/legal-ai && python scripts/sync_agents_across_companies.py --check-instructions'" >> ~/.bashrc
+source ~/.bashrc
+```
+
+- [ ] **Step 3: בדוק**
+
+```bash
+# אם Makefile:
+make -C /home/chaim/legal-ai check-agents
+# אם alias:
+check-agents
+```
+
+Expected: טבלת סוכנים מוצגת.
+
+- [ ] **Step 4: Commit (אם Makefile)**
+
+```bash
+cd /home/chaim/legal-ai
+git add Makefile
+git commit -m "feat: add check-agents and sync-agents make targets"
+```
+
+---
+
+## אימות סופי
+
+| בדיקה | פקודה | תוצאה מצופה |
+|-------|-------|-------------|
+| `--check-instructions` | `python sync_agents... --check-instructions` | טבלה עם ✅ לכל agent |
+| Pre-flight בולם apply | מחק קובץ זמנית + `--apply` | Abort עם הודעה ברורה |
+| mtime ב-DB | `SELECT metadata->>'claude_md_mtime' FROM agents` | timestamp לכל agent |
+| DRIFT זוהה | שנה קובץ + `--check-instructions` | ⚠️ DRIFT מוצג |
+| shortcut | `check-agents` או `make check-agents` | עובד |

docs/superpowers/plans/2026-05-16-scheduled-agents.md

@@ -0,0 +1,412 @@
+# Scheduled Background Agents Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** הוסף 2 cron jobs לplugin: (1) תזכורת על תיקים תקועים 3+ ימים, (2) ניתוח פידבק יו"ר שבועי עם עדכון `decision-lessons.md` אוטומטי.
+
+**Architecture:** שני jobs חדשים נרשמים ב-`ctx.jobs.register()`. Job 1 קורא `/api/cases/stale?days=3` (endpoint חדש ב-legal-ai) ומוסיף תגובה לissues תקועים. Job 2 קורא `/api/chair-feedback/weekly-summary` (endpoint חדש), שולח לCEO agent שמעדכן את `decision-lessons.md`.
+
+**Tech Stack:** TypeScript (plugin-legal-ai jobs), Python/FastAPI (legal-ai web), asyncpg, Paperclip SDK `ctx.jobs.register()`, `ctx.agents.invoke()`.
+
+---
+
+## File Map
+
+| Action | File |
+|--------|------|
+| Modify | `plugin-legal-ai/src/worker.ts` — register 2 new jobs |
+| Modify | `plugin-legal-ai/plugin.json` — declare 2 new job entries |
+| Modify | `plugin-legal-ai/src/manifest.ts` — add to jobs array |
+| Modify | `legal-ai/web/app.py` — add `GET /api/cases/stale` + `GET /api/chair-feedback/weekly-summary` |
+| Modify | `legal-ai/web/database.py` (or equivalent DB module) — add `get_stale_cases()` + `get_weekly_chair_feedback()` |
+
+---
+
+## Task 1: הוסף endpoint `GET /api/cases/stale`
+
+**Files:**
+- Modify: `legal-ai/web/app.py`
+- Modify: `legal-ai/web/database.py` (שם DB queries מנוהלים — בדוק עם `grep -n "async def get_cases\|from.*database\|import.*db" /home/chaim/legal-ai/web/app.py | head -10`)
+
+- [ ] **Step 1: מצא את module ה-DB**
+
+```bash
+grep -n "^from\|^import\|db\." /home/chaim/legal-ai/web/app.py | head -20
+```
+
+זהה את שם הmodule שמכיל את DB queries (בד"כ `database.py` או `db.py`).
+
+- [ ] **Step 2: הוסף `get_stale_cases()` לmodule ה-DB**
+
+```python
+async def get_stale_cases(days: int = 3) -> list[dict]:
+    """Return cases whose status is not 'final' and haven't been updated in `days` days."""
+    async with get_db_connection() as conn:
+        rows = await conn.fetch(
+            """
+            SELECT case_number, title, status, company_id,
+                   updated_at,
+                   now() - updated_at AS age
+            FROM cases
+            WHERE status NOT IN ('final', 'new')
+              AND updated_at < now() - ($1 || ' days')::interval
+            ORDER BY updated_at ASC
+            """,
+            str(days),
+        )
+        return [dict(r) for r in rows]
+```
+
+> `get_db_connection()` — השתמש בדפוס הקיים בקובץ. אם זה `asyncpg.connect()` ישיר, `asyncpg.create_pool()`, או context manager — העתק את הדפוס.
+
+- [ ] **Step 3: הוסף endpoint ב-`app.py`**
+
+```python
+@app.get("/api/cases/stale")
+async def list_stale_cases(days: int = 3):
+    """Cases stuck in non-final status for more than `days` days."""
+    cases = await db.get_stale_cases(days=days)
+    return {
+        "cases": [
+            {
+                "case_number": c["case_number"],
+                "title": c["title"],
+                "status": c["status"],
+                "company_id": str(c["company_id"]),
+                "days_stale": c["age"].days,
+            }
+            for c in cases
+        ],
+        "total": len(cases),
+    }
+```
+
+- [ ] **Step 4: בדיקה**
+
+```bash
+curl -s "https://legal-ai.nautilus.marcusgroup.org/api/cases/stale?days=1" | jq .total
+```
+
+Expected: JSON עם רשימת תיקים.
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add web/app.py web/database.py  # או השם הנכון
+git commit -m "feat: add GET /api/cases/stale endpoint"
+```
+
+---
+
+## Task 2: הוסף endpoint `GET /api/chair-feedback/weekly-summary`
+
+**Files:**
+- Modify: `legal-ai/web/app.py`
+- Modify: DB module
+
+- [ ] **Step 1: בדוק את מבנה טבלת `chair_feedback`**
+
+```bash
+sqlite3 /home/chaim/.paperclip/instances/default/data/app.db \
+  ".schema chair_feedback" 2>/dev/null || \
+psql -h localhost -p 5433 -U legal_ai -c "\d chair_feedback" legal_ai 2>/dev/null || \
+grep -rn "chair_feedback" /home/chaim/legal-ai/mcp-server/src/ | head -10
+```
+
+- [ ] **Step 2: הוסף `get_weekly_chair_feedback()` לDB module**
+
+```python
+async def get_weekly_chair_feedback(days: int = 7) -> list[dict]:
+    """Return chair feedback entries from the last `days` days."""
+    async with get_db_connection() as conn:
+        rows = await conn.fetch(
+            """
+            SELECT cf.case_number, cf.feedback_text, cf.created_at,
+                   cf.feedback_type, c.title
+            FROM chair_feedback cf
+            JOIN cases c ON c.case_number = cf.case_number
+            WHERE cf.created_at > now() - ($1 || ' days')::interval
+            ORDER BY cf.created_at DESC
+            """,
+            str(days),
+        )
+        return [dict(r) for r in rows]
+```
+
+> אם שמות השדות שונים (בדוק ב-Step 1) — התאם.
+
+- [ ] **Step 3: הוסף endpoint**
+
+```python
+@app.get("/api/chair-feedback/weekly-summary")
+async def get_chair_feedback_weekly(days: int = 7):
+    """Feedback entries from the past week, formatted for the learning agent."""
+    entries = await db.get_weekly_chair_feedback(days=days)
+    if not entries:
+        return {"summary": "", "entry_count": 0}
+
+    lines = [
+        f"- תיק {e['case_number']} ({e['title']}): {e['feedback_text']}"
+        for e in entries
+    ]
+    summary = "\n".join(lines)
+    return {"summary": summary, "entry_count": len(entries), "entries": entries}
+```
+
+- [ ] **Step 4: בדיקה**
+
+```bash
+curl -s "https://legal-ai.nautilus.marcusgroup.org/api/chair-feedback/weekly-summary" | jq .entry_count
+```
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd /home/chaim/legal-ai
+git add web/app.py web/database.py
+git commit -m "feat: add GET /api/chair-feedback/weekly-summary endpoint"
+```
+
+---
+
+## Task 3: הוסף jobs לplugin
+
+**Files:**
+- Modify: `plugin-legal-ai/plugin.json`
+- Modify: `plugin-legal-ai/src/manifest.ts`
+
+- [ ] **Step 1: קרא את ה-`jobs` הקיים ב-`plugin.json`**
+
+```bash
+cat /home/chaim/plugin-legal-ai/plugin.json | python3 -c "import json,sys; d=json.load(sys.stdin); print(json.dumps(d['jobs'], indent=2))"
+```
+
+- [ ] **Step 2: הוסף 2 jobs חדשים לarray `"jobs"` ב-`plugin.json`**
+
+```json
+{
+  "jobKey": "stale-case-reminder",
+  "displayName": "תזכורת תיקים תקועים",
+  "description": "מזהה תיקים שלא עודכנו 3+ ימים ומוסיף תגובה לissue",
+  "schedule": "0 8 * * *"
+},
+{
+  "jobKey": "weekly-feedback-analysis",
+  "displayName": "ניתוח פידבק שבועי",
+  "description": "מסכם פידבק יו\"ר מהשבוע האחרון ומעדכן את decision-lessons.md",
+  "schedule": "0 19 * * 0"
+}
+```
+
+> `"0 8 * * *"` = כל יום בשעה 08:00. `"0 19 * * 0"` = כל ראשון ב-19:00.
+
+- [ ] **Step 3: עדכן `manifest.ts`**
+
+```bash
+grep -n "jobs\|jobKey\|schedule" /home/chaim/plugin-legal-ai/src/manifest.ts
+```
+
+הוסף את אותם 2 objects לarray `jobs` ב-`manifest.ts`.
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /home/chaim/plugin-legal-ai
+git add plugin.json src/manifest.ts
+git commit -m "feat: declare stale-case-reminder and weekly-feedback-analysis jobs"
+```
+
+---
+
+## Task 4: Implement job handlers ב-`worker.ts`
+
+**Files:**
+- Modify: `plugin-legal-ai/src/worker.ts`
+
+- [ ] **Step 1: קרא את handler של `sync-case-status` הקיים**
+
+```bash
+grep -n "sync-case-status\|jobs.register\|jobKey" /home/chaim/plugin-legal-ai/src/worker.ts
+```
+
+העתק את הדפוס.
+
+- [ ] **Step 2: הוסף את `stale-case-reminder` handler**
+
+בתוך `setup(ctx)`, אחרי רישום ה-job הקיים:
+
+```typescript
+ctx.jobs.register("stale-case-reminder", async (job) => {
+  ctx.logger.info("stale-case-reminder: starting");
+
+  const config = await ctx.config.get();
+  const apiBase = (config.legalApiBaseUrl as string) ?? "http://localhost:8085";
+
+  const resp = await ctx.http.fetch(`${apiBase}/api/cases/stale?days=3`);
+  if (!resp.ok) {
+    ctx.logger.error(`stale-case-reminder: API error ${resp.status}`);
+    return;
+  }
+
+  const data = (await resp.json()) as {
+    cases: Array<{
+      case_number: string;
+      title: string;
+      status: string;
+      company_id: string;
+      days_stale: number;
+    }>;
+  };
+
+  for (const staleCase of data.cases) {
+    const issueId = await ctx.state.get(
+      { companyId: staleCase.company_id },
+      `case:${staleCase.case_number}`
+    );
+    if (!issueId) continue;
+
+    await ctx.issues.createComment({
+      issueId: issueId as string,
+      body: `⚠️ **תיק תקוע** — ${staleCase.days_stale} ימים ללא עדכון (סטטוס: ${staleCase.status}). האם יש צורך בפעולה?`,
+    });
+
+    ctx.logger.info(
+      `stale-case-reminder: posted reminder for ${staleCase.case_number} (${staleCase.days_stale}d stale)`
+    );
+  }
+
+  ctx.logger.info(`stale-case-reminder: done. ${data.cases.length} cases reminded`);
+});
+```
+
+- [ ] **Step 3: הוסף את `weekly-feedback-analysis` handler**
+
+```typescript
+ctx.jobs.register("weekly-feedback-analysis", async (job) => {
+  ctx.logger.info("weekly-feedback-analysis: starting");
+
+  const config = await ctx.config.get();
+  const apiBase = (config.legalApiBaseUrl as string) ?? "http://localhost:8085";
+
+  const resp = await ctx.http.fetch(`${apiBase}/api/chair-feedback/weekly-summary`);
+  if (!resp.ok) {
+    ctx.logger.error(`weekly-feedback-analysis: API error ${resp.status}`);
+    return;
+  }
+
+  const data = (await resp.json()) as {
+    summary: string;
+    entry_count: number;
+  };
+
+  if (data.entry_count === 0) {
+    ctx.logger.info("weekly-feedback-analysis: no feedback this week, skipping");
+    return;
+  }
+
+  // Invoke the CEO agent to process feedback and update decision-lessons.md
+  const companies = await ctx.companies.list();
+  for (const company of companies) {
+    // CEO IDs per company
+    const CEO_IDS: Record<string, string> = {
+      "42a7acd0-30c5-4cbd-ac97-7424f65df294": "752cebdd-6748-4a04-aacd-c7ab0294ef33",
+      "8639e837-4c9d-47fa-a76b-95788d651896": "cdbfa8bc-3d61-41a4-a2e7-677ec7d34562",
+    };
+    const ceoId = CEO_IDS[company.id];
+    if (!ceoId) continue;
+
+    await ctx.agents.invoke(ceoId, company.id, {
+      prompt: `ניתוח פידבק שבועי יו"ר (${data.entry_count} פריטים):
+
+${data.summary}
+
+המשימה: עדכן את /home/chaim/legal-ai/docs/legal-decision-lessons.md עם הלקחים החדשים שעולים מהפידבק. הוסף רק לקחים חדשים שלא קיימים כבר. קבץ לפי נושא.`,
+      reason: "weekly-feedback-analysis scheduled job",
+    });
+
+    ctx.logger.info(
+      `weekly-feedback-analysis: invoked CEO for company ${company.id} (${data.entry_count} feedback entries)`
+    );
+    break; // One CEO is enough — lessons file is shared
+  }
+});
+```
+
+- [ ] **Step 4: TypeScript check**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npx tsc --noEmit
+```
+
+Expected: 0 errors.
+
+- [ ] **Step 5: Build**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npm run build
+```
+
+- [ ] **Step 6: Commit**
+
+```bash
+cd /home/chaim/plugin-legal-ai
+git add src/worker.ts
+git commit -m "feat: implement stale-case-reminder and weekly-feedback-analysis jobs"
+```
+
+---
+
+## Task 5: Re-install plugin + בדיקה
+
+- [ ] **Step 1: Deploy legal-ai**
+
+```bash
+cd /home/chaim/legal-ai && git push origin main
+# המתן ~3 דקות
+curl -s https://legal-ai.nautilus.marcusgroup.org/api/health | jq .status
+```
+
+- [ ] **Step 2: Re-install plugin**
+
+```bash
+cd /home/chaim/plugin-legal-ai && npm run build
+npx paperclipai plugin uninstall marcusgroup.legal-ai \
+  --api-base http://localhost:3100 --api-key pcapi_legal_install_key_2026
+npx paperclipai plugin install /home/chaim/plugin-legal-ai \
+  --api-base http://localhost:3100 --api-key pcapi_legal_install_key_2026
+pm2 restart paperclip
+```
+
+- [ ] **Step 3: בדוק שה-jobs רשומים**
+
+```bash
+curl -s -H "Authorization: Bearer pcapi_legal_install_key_2026" \
+  http://localhost:3100/api/plugins/marcusgroup.legal-ai/jobs | jq .[].jobKey
+```
+
+Expected: `"sync-case-status"`, `"stale-case-reminder"`, `"weekly-feedback-analysis"`.
+
+- [ ] **Step 4: הפעל job ידנית לבדיקה**
+
+```bash
+curl -s -X POST -H "Authorization: Bearer pcapi_legal_install_key_2026" \
+  http://localhost:3100/api/plugins/marcusgroup.legal-ai/jobs/stale-case-reminder/run | jq .
+```
+
+Expected: Job הופעל. בדוק logs:
+```bash
+pm2 logs paperclip --lines 30 | grep "stale-case-reminder"
+```
+
+---
+
+## אימות סופי
+
+| בדיקה | פקודה | תוצאה מצופה |
+|-------|-------|-------------|
+| API stale endpoint | `curl .../api/cases/stale?days=1` | JSON עם cases |
+| API feedback endpoint | `curl .../api/chair-feedback/weekly-summary` | JSON עם summary |
+| Jobs רשומים | `GET .../api/plugins/.../jobs` | 3 jobs רשומים |
+| Stale reminder ידני | `POST .../jobs/stale-case-reminder/run` | תגובות בissues |
+| Feedback analysis ידני | `POST .../jobs/weekly-feedback-analysis/run` | CEO מועיר |

mcp-server/src/legal_mcp/config.py

@@ -275,8 +275,8 @@ HALACHA_CANONICAL_SYNTH_MODEL = os.environ.get("HALACHA_CANONICAL_SYNTH_MODEL",
 HALACHA_CANONICAL_SYNTH_EFFORT = os.environ.get("HALACHA_CANONICAL_SYNTH_EFFORT", "high")
 HALACHA_CANONICAL_SYNTH_DRIFT_FLOOR = float(os.environ.get("HALACHA_CANONICAL_SYNTH_DRIFT_FLOOR", "0.80"))
 
-# Google Cloud Vision (OCR for scanned PDFs)
+# Mistral OCR (fallback for scanned PDFs — replaces Google Cloud Vision)
-GOOGLE_CLOUD_VISION_API_KEY = os.environ.get("GOOGLE_CLOUD_VISION_API_KEY", "")
+MISTRAL_API_KEY = os.environ.get("MISTRAL_API_KEY", "")
 
 # Data directory
 DATA_DIR = Path(os.environ.get("DATA_DIR", str(Path.home() / "legal-ai" / "data")))
@@ -361,7 +361,7 @@ PARENT_DOC_CHILD_OVERLAP_TOKENS = int(
 # External service allowlist — case materials may ONLY be sent to these domains
 ALLOWED_EXTERNAL_SERVICES = {
     "api.voyageai.com",         # Voyage AI (embeddings)
-    "vision.googleapis.com",    # Google Cloud Vision (OCR)
+    "api.mistral.ai",           # Mistral OCR (scanned PDFs)
 }
 
 # Audit

mcp-server/src/legal_mcp/server.py

@@ -418,6 +418,12 @@ async def digest_process_pending(limit: int = 20) -> str:
     return await digest_tools.digest_process_pending(_clamp_limit(limit))
 
 
+@mcp.tool()
+async def digest_radar(case_number: str, limit: int = 5, min_score: float = 0.45) -> str:
+    """רדאר-יומונים הקשרי-לתיק (X12) — מחזיר יומונים לא-מקושרים (פס"ד שעוד אין לנו) שהנושא שלהם קרוב סמנטית לתיק, כדי שלא ייפול פס"ד רלוונטי בזמן ההכרעה. כל ליד מצביע על מראה-המקום של הפס"ד המקורי + סטטוס-הפער + פעולה מוצעת. ⚠️ radar בלבד — היומון אינו מצוטט (INV-DIG1)."""
+    return await digest_tools.digest_radar(case_number, _clamp_limit(limit), min_score)
+
+
 @mcp.tool()
 async def halacha_review(
     halacha_id: str,
@@ -1170,6 +1176,15 @@ async def list_chair_feedback(
     return await workflow.list_chair_feedback(case_number, category, unresolved_only, _clamp_limit(limit))
 
 
+@mcp.tool()
+async def record_curator_findings(case_number: str, findings: list[dict]) -> str:
+    """לכידת ממצאי-האוצֵר על החלטה סופית כ-decision_lessons מובְנים (source='curator',
+    ממתינים לשער-יו"ר) — INV-LRN3. כל ממצא: {"text": "...", "category": style/structure/
+    lexicon/tabular} (או "tag" עברי). מחזיר כמה נרשמו וכמה כפילויות דולגו. read-only על
+    התוכן — הרישום הצעה הממתינה לאישור דפנה ב-/training, לא שינוי-קול ישיר (G10)."""
+    return await workflow.record_curator_findings(case_number, findings)
+
+
 @mcp.tool()
 async def halacha_corroboration(halacha_id: str) -> dict:
     """החזר את ה-corroboration של הלכה: הציטוטים שמתקפים אותה, הטיפול, וסיכום (X11, read-only)."""

mcp-server/src/legal_mcp/services/argument_aggregator.py

@@ -230,6 +230,15 @@ async def aggregate_claims_to_arguments(
             party = "unknown"
         by_party.setdefault(party, []).append(dict(r))
 
+    # Valid claim_ids for this case == the ids of the claims we just fetched.
+    # The LLM is asked to echo back supporting claim_ids, but it may hallucinate
+    # a syntactically-valid-but-nonexistent UUID (malformed ones are already
+    # dropped in ``_normalize_argument``). Validating against this known set at
+    # source keeps a doomed INSERT — which would poison the surrounding asyncpg
+    # transaction (FK violation -> "current transaction is aborted") — out of
+    # the transaction entirely (G1: fix at source, not symptom).
+    valid_claim_ids: set[UUID] = {r["id"] for r in rows}
+
     party_counts: dict[str, int] = {}
     inserted = 0
     errors: list[str] = []
@@ -275,17 +284,30 @@ async def aggregate_claims_to_arguments(
                         arg["priority"],
                     )
                     for cid in arg["claim_ids"]:
-                        try:
+                        if cid not in valid_claim_ids:
-                            await conn.execute(
+                            # Hallucinated claim_id that doesn't belong to this
-                                """INSERT INTO legal_argument_propositions
+                            # case. Skip it rather than letting the FK violation
-                                    (argument_id, claim_id)
+                            # abort the whole transaction.
-                                   VALUES ($1, $2)
+                            logger.warning(
-                                   ON CONFLICT DO NOTHING""",
+                                "argument_aggregator: skipped unknown claim_id %s for arg %s",
-                                arg_id, cid,
+                                cid, arg_id,
                             )
+                            continue
+                        try:
+                            # Per-row savepoint: even after the validation above,
+                            # wrap the INSERT so any unexpected constraint failure
+                            # rolls back to the savepoint instead of poisoning the
+                            # surrounding transaction (asyncpg nests transaction()
+                            # as SAVEPOINT when already inside one).
+                            async with conn.transaction():
+                                await conn.execute(
+                                    """INSERT INTO legal_argument_propositions
+                                        (argument_id, claim_id)
+                                       VALUES ($1, $2)
+                                       ON CONFLICT DO NOTHING""",
+                                    arg_id, cid,
+                                )
                         except Exception as e:  # noqa: BLE001
-                            # Likely FK violation if the LLM hallucinated
-                            # a claim_id. Log and continue.
                             logger.warning(
                                 "argument_aggregator: skipped bad claim_id %s for arg %s: %s",
                                 cid, arg_id, e,

mcp-server/src/legal_mcp/services/block_writer.py

@@ -983,6 +983,22 @@ async def _build_style_context(practice_area: str = "") -> str:
             ("anti_patterns", "אנטי-דפוסים (להימנע)"),
         ):
             ov = await db.get_methodology_overrides(cat)
+            if cat == "anti_patterns":
+                # Anti-patterns are STRUCTURAL INVARIANTS of Dafna's style (no
+                # markdown headers, no bullet lists, no mid-paragraph mini-lists —
+                # she writes continuous legal narrative). They must reach the writer
+                # ALWAYS, from the SAME canonical list style_distance measures against
+                # (lessons.ANTI_PATTERNS) — otherwise the loop detects them but never
+                # corrects them, and drafts keep emitting them (the gap that left
+                # 8137 with 28 hits). Chair additions layer on top; they never
+                # remove the canonical ones.
+                from legal_mcp.services.lessons import ANTI_PATTERNS as _ANTI
+                learned.append(f"\n**{label} — כתוב נרטיב משפטי רציף; הימנע מ:**")
+                for ap in _ANTI:
+                    learned.append(f"- {ap['note']}")
+                for k, v in (ov or {}).items():
+                    learned.append(f"- (יו\"ר) {k}: {json.dumps(v, ensure_ascii=False)}")
+                continue
             if ov:
                 learned.append(f"\n**{label} — ערכי היו\"ר (גוברים על ברירת-המחדל):**")
                 for k, v in ov.items():

mcp-server/src/legal_mcp/services/case_citation_verification.py

@@ -0,0 +1,134 @@
+"""Citation-verification view (X11 Phase 2 / #154) — the chair's "אימות פסיקה" tab.
+
+Assembles, per legal ARGUMENT of a case, the supporting precedents the chair should
+verify before the writer cites them:
+
+  • in-corpus suggestions — per-issue semantic retrieval over the authoritative
+    precedent library (``search_library``), each carrying the cumulative authority
+    signal (``cited_by``: followed/distinguished — db.citation_authority, X11).
+  • attached/verified state — any ``case_precedents`` row already attached to the
+    argument (verified flag + chair_note), merged onto the matching suggestion.
+  • radar — UNLINKED digests relevant to the same issue (rulings we don't hold yet),
+    from ``case_digest_radar`` grouped by matched issue.
+
+Pure read/assembly — never writes, never cites (INV-DIG1/INV-AH). The chair verifies
+through ``db.set_case_precedent_verified`` / attach; the writer consumes only verified
+rows. Reuses the one corpus search + the one authority query + the one radar (G2).
+"""
+
+from __future__ import annotations
+
+import logging
+from uuid import UUID
+
+from legal_mcp.services import (
+    argument_aggregator,
+    db,
+    digest_library,
+    precedent_library,
+)
+
+logger = logging.getLogger(__name__)
+
+_SUGGEST_PER_ISSUE = 4
+_SUGGEST_FLOOR = 0.45
+
+
+async def build_view(case_number: str) -> dict:
+    case = await db.get_case_by_number(case_number)
+    if not case:
+        return {"status": "case_not_found", "case_number": case_number, "arguments": []}
+    case_id = case["id"]
+    if isinstance(case_id, str):
+        case_id = UUID(case_id)
+    ctx = " ".join(x for x in [case.get("title") or "", case.get("appeal_subtype") or ""] if x).strip()
+
+    args = await argument_aggregator.get_legal_arguments(case_id)
+
+    # Attached precedents already on the case → grouped by argument_id, keyed by
+    # the resolved corpus ruling so we can merge verify-state onto a suggestion.
+    attached = await db.list_case_precedents(case_id)
+    attached_by_arg: dict[str, dict[str, dict]] = {}
+    for p in attached:
+        aid = str(p.get("argument_id") or "")
+        clid = str(p.get("case_law_id") or "")
+        if aid and clid:
+            attached_by_arg.setdefault(aid, {})[clid] = p
+
+    # Radar (unlinked digests) once, grouped by the issue label it matched.
+    radar_by_issue: dict[str, list[dict]] = {}
+    try:
+        radar = await digest_library.case_digest_radar(case_number, limit=12, min_score=0.42)
+        for lead in radar.get("leads", []):
+            for label in (lead.get("matched_issues") or ["—"]):
+                radar_by_issue.setdefault(label, []).append(lead)
+    except Exception as e:  # noqa: BLE001 — radar is best-effort
+        logger.warning("citation_verification radar failed for %s: %s", case_number, e)
+
+    out_args: list[dict] = []
+    n_verified = 0
+    for a in args:
+        aid = str(a["id"])
+        title = (a.get("argument_title") or "").strip()
+        topic = (a.get("legal_topic") or "").strip()
+        query = f"{ctx} {title}. {topic}".strip()
+
+        hits = []
+        try:
+            hits = await precedent_library.search_library(
+                query=query, limit=_SUGGEST_PER_ISSUE, include_halachot=True)
+        except Exception as e:  # noqa: BLE001
+            logger.warning("citation_verification search failed (%s): %s", title[:30], e)
+
+        # Resolve the authority breakdown for the hit set in one batched query.
+        clids = [UUID(str(h["case_law_id"])) for h in hits
+                 if h.get("case_law_id") and float(h.get("score", 0) or 0) >= _SUGGEST_FLOOR]
+        authority = await db.citation_authority(clids) if clids else {}
+
+        seen: set[str] = set()
+        supporting: list[dict] = []
+        for h in hits:
+            clid = str(h.get("case_law_id") or "")
+            if not clid or clid in seen:
+                continue
+            if float(h.get("score", 0) or 0) < _SUGGEST_FLOOR:
+                continue
+            seen.add(clid)
+            att = attached_by_arg.get(aid, {}).get(clid)
+            if att and att.get("verified"):
+                n_verified += 1
+            supporting.append({
+                "case_law_id": clid,
+                "case_number": h.get("case_number") or "",
+                "case_name": h.get("case_name") or "",
+                "quote": h.get("supporting_quote") or h.get("rule_statement") or "",
+                "score": round(float(h.get("score", 0) or 0), 3),
+                "cited_by": authority.get(clid, {"total": 0, "positive": 0,
+                                                 "negative": 0, "unclassified": 0,
+                                                 "by_treatment": {}}),
+                "attached_id": str(att["id"]) if att else None,
+                "verified": bool(att.get("verified")) if att else False,
+                "chair_note": (att.get("chair_note") or "") if att else "",
+            })
+
+        out_args.append({
+            "argument_id": aid,
+            "title": title,
+            "legal_topic": topic,
+            "priority": a.get("priority") or "",
+            "party": a.get("party") or "",
+            "supporting": supporting,
+            "radar": radar_by_issue.get(title, []),
+        })
+
+    return {
+        "status": "ok",
+        "case_number": case_number,
+        "arguments": out_args,
+        "summary": {
+            "arguments_total": len(out_args),
+            "arguments_with_support": sum(1 for x in out_args if x["supporting"]),
+            "verified": n_verified,
+            "radar_leads": sum(len(x["radar"]) for x in out_args),
+        },
+    }

mcp-server/src/legal_mcp/services/chunker.py

@@ -144,9 +144,12 @@ def _split_into_sections(text: str) -> list[tuple[str, str]]:
     markers: list[tuple[int, str]] = []
 
     for pattern, section_type in SECTION_PATTERNS:
-        # ^ + MULTILINE: line start only. Optional leading spaces/tabs and an
+        # ^ + MULTILINE: line start only. Optional leading spaces/tabs, an
+        # optional Markdown ATX header prefix (``## ``/``### ``), and an
         # optional ordinal prefix ("5.", "5)", "ג.") before the keyword.
-        anchored = rf"^[ \t]*(?:\d+[.)]\s*|[א-ת][.)]\s*)?(?:{pattern})"
+        # The Markdown prefix handles Mistral OCR output where section
+        # titles are rendered as ``## נימוקי הערר`` etc.
+        anchored = rf"^[ \t]*(?:#{1,3}\s+)?(?:\d+[.)]\s*|[א-ת][.)]\s*)?(?:{pattern})"
         for match in re.finditer(anchored, text, re.MULTILINE):
             markers.append((match.start(), section_type))
 

mcp-server/src/legal_mcp/services/citation_extractor.py

@@ -236,7 +236,18 @@ async def extract_and_store(case_law_id: UUID) -> dict:
     if not row:
         return {"extracted": 0, "linked": 0, "new": 0, "skipped": 0, "error": "not_found"}
 
-    text = row["full_text"] or ""
+    # A citation counts as the DECIDING body's reliance ONLY when it appears in
+    # the discussion/ruling sections — never where a party cites a ruling in its
+    # own arguments (block ז). Reuse the halacha extractor's section selection
+    # (G2: a single definition of "reasoning sections" — legal_analysis/ruling/
+    # conclusion + the discussion-anchor, excluding facts/claims/intro) so a
+    # ruling cited only in appellant_claims/parties_claims is never attributed to
+    # the chair. Fall back to full_text only for un-chunked rows.
+    from legal_mcp.services import halacha_extractor as _hx
+    disc_chunks, _used_fallback = await _hx._select_extractable_chunks(case_law_id)
+    text = "\n".join((c.get("content") or "") for c in disc_chunks).strip()
+    if not text:
+        text = row["full_text"] or ""
     own_norm = _normalize_case_number(row["case_number"] or "")
 
     extracted = 0
@@ -388,12 +399,16 @@ async def list_citations_to_case_law(case_law_id: UUID) -> list[dict]:
                pic.cited_case_number,
                pic.match_context,
                pic.match_pattern,
+               pic.treatment,
                pic.confidence::float AS confidence,
                pic.created_at,
                cl.case_number AS source_case_number,
                cl.case_name AS source_case_name,
                cl.chair_name AS source_chair_name,
-               cl.district AS source_district
+               cl.district AS source_district,
+               cl.precedent_level AS source_precedent_level,
+               cl.court AS source_court,
+               cl.date AS source_date
           FROM precedent_internal_citations pic
           JOIN case_law cl ON cl.id = pic.source_case_law_id
          WHERE pic.cited_case_law_id = $1
@@ -432,3 +447,42 @@ async def get_cited_case_law_ids(source_case_law_ids: list[UUID]) -> dict[str, l
     for r in rows:
         out.setdefault(r["source_id"], []).append(r["cited_id"])
     return out
+
+
+async def relink_orphan_citations(case_law_id: "UUID | None" = None) -> int:
+    """Back-fill ``cited_case_law_id`` on orphan citation edges that now resolve.
+
+    Citations are resolved to a corpus row only at extraction time (see
+    ``extract_and_store``). When a cited ruling is uploaded LATER, its existing
+    orphan edges (``cited_case_law_id IS NULL``) are never re-resolved, so the
+    citation graph under-counts real connectivity. Call on every precedent
+    upload (``case_law_id`` set → link only edges that resolve to that row) or as
+    a one-off sweep (``case_law_id=None`` → re-resolve every orphan). Reuses the
+    canonical resolver ``_resolve_case_law_id`` (no parallel matching logic, G2);
+    idempotent — a second run links nothing new.
+    """
+    pool = await db.get_pool()
+    async with pool.acquire() as conn:
+        orphans = await conn.fetch(
+            "SELECT DISTINCT cited_case_number FROM precedent_internal_citations "
+            "WHERE cited_case_law_id IS NULL AND coalesce(cited_case_number, '') <> ''"
+        )
+    linked = 0
+    for r in orphans:
+        cited = await _resolve_case_law_id(_normalize_case_number(r["cited_case_number"]))
+        if cited is None:
+            continue
+        if case_law_id is not None and cited != case_law_id:
+            continue
+        async with pool.acquire() as conn:
+            res = await conn.execute(
+                "UPDATE precedent_internal_citations "
+                "SET cited_case_law_id = $1, confidence = GREATEST(confidence, 0.90) "
+                "WHERE cited_case_law_id IS NULL AND cited_case_number = $2",
+                cited, r["cited_case_number"],
+            )
+        try:
+            linked += int(res.split()[-1])
+        except (ValueError, IndexError):
+            pass
+    return linked

mcp-server/src/legal_mcp/services/db.py

@@ -1697,6 +1697,80 @@ CREATE INDEX IF NOT EXISTS idx_hcc_canonical
 """
 
 
+SCHEMA_V42_SQL = """
+-- INV-DM7: authority (binding/persuasive) is STRUCTURAL for committee sources.
+-- An appeals-committee decision (source_kind='internal_committee', or any row
+-- with source_type='appeals_committee') is persuasive, NEVER binding — a
+-- committee's ruling does not bind another committee, nor itself. is_binding
+-- stays a stored column (it selects the halacha-extraction prompt), but for
+-- committee sources it is a faithful cache of the derived value, not a chair
+-- guess. Mirrors 02-data-model.md §INV-DM7 / X8-field-provenance.md.
+-- Step 1 — normalize legacy rows at the source (G1), idempotent.
+UPDATE case_law SET is_binding = FALSE
+    WHERE is_binding IS TRUE
+      AND (source_kind = 'internal_committee' OR source_type = 'appeals_committee');
+-- Step 2 — constrain so a binding committee row can never be written again.
+DO $$ BEGIN
+    ALTER TABLE case_law ADD CONSTRAINT case_law_committee_not_binding_check
+        CHECK (NOT ((source_kind = 'internal_committee'
+                     OR source_type = 'appeals_committee') AND is_binding));
+EXCEPTION WHEN duplicate_object THEN NULL; END $$;
+"""
+
+
+SCHEMA_V43_SQL = """
+-- חיפה (Haifa) is a distinct planning district, separate from הצפון (North).
+-- A bug in _district_from_court mapped "חיפה" -> "צפון", and the service-layer
+-- _VALID_DISTRICTS omitted "חיפה" entirely, so Haifa committee decisions were
+-- mis-filed under צפון. Both fixed in internal_decisions.py; this reclassifies
+-- the legacy rows at the source (G1) by their court text. Idempotent.
+UPDATE case_law SET district = 'חיפה'
+    WHERE source_kind = 'internal_committee'
+      AND district = 'צפון'
+      AND court ILIKE '%חיפה%';
+"""
+
+
+SCHEMA_V44_SQL = """
+-- Citation-verification panel (#154): the chair verifies each corpus precedent
+-- against the specific legal ARGUMENT before the writer may cite it (INV-AH gate).
+-- case_precedents gains: the argument it supports, the resolved case_law row
+-- (for the cited_by authority signal + dedup), a verified flag (the gate), and a
+-- verification timestamp. chair_note already exists. All nullable so legacy
+-- section-scoped rows stay valid.
+ALTER TABLE case_precedents ADD COLUMN IF NOT EXISTS argument_id UUID
+    REFERENCES legal_arguments(id) ON DELETE SET NULL;
+ALTER TABLE case_precedents ADD COLUMN IF NOT EXISTS case_law_id UUID
+    REFERENCES case_law(id) ON DELETE SET NULL;
+ALTER TABLE case_precedents ADD COLUMN IF NOT EXISTS verified BOOLEAN DEFAULT false;
+ALTER TABLE case_precedents ADD COLUMN IF NOT EXISTS verified_at TIMESTAMPTZ;
+CREATE INDEX IF NOT EXISTS idx_case_precedents_argument ON case_precedents(argument_id);
+"""
+
+
+# V45 (Path A — prospective held-out): a style-distance snapshot captured at
+# final-upload time, BEFORE this case's lessons are folded. Because the draft was
+# written with only the PRIOR lesson pool, each row is a clean generalization
+# datapoint: "with N accumulated lessons, our draft on this unseen case scored X".
+# pool_* record how much voice-learning had accumulated when the draft was made,
+# so the trend (style-distance vs pool size / time) shows whether learning
+# generalizes. Append-only; one row per final upload. (07-learning §0, INV-LRN4.)
+SCHEMA_V45_SQL = """
+CREATE TABLE IF NOT EXISTS style_distance_history (
+    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
+    case_number TEXT NOT NULL,
+    pair_id UUID REFERENCES draft_final_pairs(id) ON DELETE SET NULL,
+    measured_at TIMESTAMPTZ DEFAULT now(),
+    pool_discussion_rules INT DEFAULT 0,
+    pool_transition_phrases INT DEFAULT 0,
+    anti_pattern_total INT,
+    ratio_max_deviation REAL,
+    change_percent REAL
+);
+CREATE INDEX IF NOT EXISTS idx_style_distance_history_measured ON style_distance_history(measured_at);
+"""
+
+
 # Stable, arbitrary key for the session-level advisory lock that serialises
 # schema DDL across processes. Every short-lived process (cron drains, services)
 # re-runs the idempotent migrations on startup; without this lock two processes
@@ -1714,7 +1788,7 @@ async def _run_schema_migrations(pool: asyncpg.Pool) -> None:
             await _apply_schema_ddl(conn)
         finally:
             await conn.execute("SELECT pg_advisory_unlock($1)", _MIGRATION_LOCK_KEY)
-    logger.info("Database schema initialized (v1-v41)")
+    logger.info("Database schema initialized (v1-v43)")
 
 
 async def _apply_schema_ddl(conn: asyncpg.Connection) -> None:
@@ -1760,6 +1834,10 @@ async def _apply_schema_ddl(conn: asyncpg.Connection) -> None:
         await conn.execute(SCHEMA_V39_SQL)
         await conn.execute(SCHEMA_V40_SQL)
         await conn.execute(SCHEMA_V41_SQL)
+        await conn.execute(SCHEMA_V42_SQL)
+        await conn.execute(SCHEMA_V43_SQL)
+        await conn.execute(SCHEMA_V44_SQL)
+        await conn.execute(SCHEMA_V45_SQL)
 
 
 async def init_schema() -> None:
@@ -2597,6 +2675,22 @@ async def add_decision_lesson(
     return dict(row) if row else {}
 
 
+async def get_style_corpus_id_by_decision(decision_number: str) -> UUID | None:
+    """Resolve the style_corpus row id for a final decision by its number.
+
+    The curator knows the case_number; lessons attach to the corpus row the
+    learning pipeline enrolled (enroll_style_corpus). Returns None if the final
+    has not been enrolled yet (caller surfaces it — no silent attach).
+    """
+    pool = await get_pool()
+    async with pool.acquire() as conn:
+        row = await conn.fetchrow(
+            "SELECT id FROM style_corpus WHERE decision_number = $1 LIMIT 1",
+            decision_number,
+        )
+    return row["id"] if row else None
+
+
 async def update_decision_lesson(
     lesson_id: UUID,
     *,
@@ -2781,6 +2875,48 @@ async def get_style_patterns(pattern_type: str | None = None) -> list[dict]:
     return [dict(r) for r in rows]
 
 
+async def append_global_rule(
+    category: str, key: str, items: list[str], seed_if_missing: list | None = None,
+) -> int:
+    """Append items to a _global appeal_type_rules list value (the writer-consumed
+    methodology channel), idempotently and under a row lock. Returns how many NEW
+    items were added (existing duplicates are skipped). Single locked read-modify-
+    write (MET-2/3) so concurrent appends can't drop items. Shared by the /training
+    promote gate (web `_append_methodology_override`) and chair-feedback auto-flow
+    (G2 — one append implementation)."""
+    pool = await get_pool()
+    async with pool.acquire() as conn:
+        async with conn.transaction():
+            row = await conn.fetchrow(
+                "SELECT rule_value FROM appeal_type_rules "
+                "WHERE appeal_type = '_global' AND rule_category = $1 AND rule_key = $2 "
+                "FOR UPDATE",
+                category, key,
+            )
+            if row:
+                current = row["rule_value"]
+                if isinstance(current, str):
+                    try:
+                        current = json.loads(current)
+                    except (json.JSONDecodeError, TypeError):
+                        current = []
+            else:
+                current = list(seed_if_missing or [])
+            if not isinstance(current, list):
+                current = []
+            added = [s for s in items if s and s not in current]
+            if not added and row:
+                return 0
+            merged = current + added
+            await conn.execute(
+                "INSERT INTO appeal_type_rules (id, appeal_type, rule_category, rule_key, rule_value) "
+                "VALUES (gen_random_uuid(), '_global', $1, $2, $3::text::jsonb) "
+                "ON CONFLICT (appeal_type, rule_category, rule_key) DO UPDATE SET rule_value = $3::text::jsonb",
+                category, key, json.dumps(merged, ensure_ascii=False),
+            )
+            return len(added)
+
+
 async def get_methodology_overrides(category: str) -> dict:
     """Chair's /methodology edits for one category (golden_ratios / discussion_rules /
     content_checklists). Returns {rule_key: parsed_value}. These OVERRIDE the hardcoded
@@ -2805,6 +2941,66 @@ async def get_methodology_overrides(category: str) -> dict:
     return out
 
 
+async def voice_lesson_pool_sizes() -> dict:
+    """How many folded voice-learning items are in the writer-consumed pool right now
+    (universal discussion_rules + transition_phrases). Used by the prospective held-out
+    snapshot to record how much learning had accumulated when a draft was produced."""
+    pool = await get_pool()
+    async with pool.acquire() as conn:
+        rows = await conn.fetch(
+            "SELECT rule_category, COALESCE(jsonb_array_length(rule_value), 0) AS n "
+            "FROM appeal_type_rules "
+            "WHERE appeal_type = '_global' AND rule_key = 'universal' "
+            "AND rule_category IN ('discussion_rules', 'transition_phrases')",
+        )
+    out = {"discussion_rules": 0, "transition_phrases": 0}
+    for r in rows:
+        out[r["rule_category"]] = r["n"]
+    return out
+
+
+async def record_style_distance_snapshot(
+    case_number: str, pair_id: str | None, pool_rules: int, pool_phrases: int,
+    anti_pattern_total: int | None, ratio_max_deviation: float | None,
+    change_percent: float | None,
+) -> None:
+    """Append one prospective held-out datapoint (Path A). Append-only; never updates."""
+    pool = await get_pool()
+    async with pool.acquire() as conn:
+        await conn.execute(
+            "INSERT INTO style_distance_history (case_number, pair_id, "
+            "pool_discussion_rules, pool_transition_phrases, anti_pattern_total, "
+            "ratio_max_deviation, change_percent) VALUES ($1, $2, $3, $4, $5, $6, $7)",
+            case_number, UUID(pair_id) if pair_id else None,
+            pool_rules, pool_phrases, anti_pattern_total,
+            ratio_max_deviation, change_percent,
+        )
+
+
+async def get_style_distance_history() -> list[dict]:
+    """The prospective held-out trend, oldest-first (Path A). Each row = one final
+    upload's style-distance measured before that case's lessons were folded."""
+    pool = await get_pool()
+    async with pool.acquire() as conn:
+        rows = await conn.fetch(
+            "SELECT case_number, measured_at, pool_discussion_rules, "
+            "pool_transition_phrases, anti_pattern_total, ratio_max_deviation, "
+            "change_percent FROM style_distance_history ORDER BY measured_at",
+        )
+    return [
+        {
+            "case_number": r["case_number"],
+            "measured_at": r["measured_at"].isoformat() if r["measured_at"] else None,
+            "pool_discussion_rules": r["pool_discussion_rules"],
+            "pool_transition_phrases": r["pool_transition_phrases"],
+            "anti_pattern_total": r["anti_pattern_total"],
+            "ratio_max_deviation": r["ratio_max_deviation"],
+            "change_percent": r["change_percent"],
+        }
+        for r in rows
+    ]
+
+
 async def get_recent_decision_lessons(limit: int = 15, practice_area: str = "") -> list[dict]:
     """Per-decision learnings the chair/curator attached in /training (decision_lessons),
     so the writer consumes them too (T15). Prefers style/structure/lexicon, recent first.
@@ -2812,6 +3008,12 @@ async def get_recent_decision_lessons(limit: int = 15, practice_area: str = "")
     Gate (INV-LRN1/G10): returns only CHAIR-APPROVED lessons (review_status='approved').
     Panel-written proposals stay out of the writer's context until the chair approves
     them in /training — the chair's review is the gate, not the panel's 2/2 vote.
+
+    No silent cap (חוקה §6): when more approved lessons match than `limit`, the
+    overflow is dropped by ORDER BY created_at DESC — log it so coverage loss is
+    observable, not invisible. The real fix is lesson synthesis (TaskMaster #158):
+    consolidate overlapping lessons into fewer, richer ones so the set fits without
+    truncation. Until then this WARN is the audit trail.
     """
     pool = await get_pool()
     async with pool.acquire() as conn:
@@ -2826,6 +3028,20 @@ async def get_recent_decision_lessons(limit: int = 15, practice_area: str = "")
                LIMIT $1""",
             limit, practice_area,
         )
+        if len(rows) >= limit:
+            total = await conn.fetchval(
+                """SELECT count(*) FROM decision_lessons dl
+                   JOIN style_corpus sc ON sc.id = dl.style_corpus_id
+                   WHERE dl.review_status = 'approved'
+                     AND ($1 = '' OR sc.practice_area = $1)""",
+                practice_area,
+            ) or 0
+            if total > limit:
+                logger.warning(
+                    "get_recent_decision_lessons: capped %d→%d approved lessons "
+                    "(practice_area=%r) — %d not reaching the writer. See TaskMaster #158 (synthesis).",
+                    total, limit, practice_area or "*", total - limit,
+                )
     return [dict(r) for r in rows]
 
 
@@ -3295,28 +3511,58 @@ async def create_case_precedent(
     chair_note: str = "",
     pdf_document_id: UUID | None = None,
     practice_area: str | None = None,
+    argument_id: UUID | None = None,
+    case_law_id: UUID | None = None,
+    verified: bool = False,
 ) -> dict:
-    """Insert a new precedent attached to a case."""
+    """Insert a new precedent attached to a case.
+
+    ``argument_id``/``case_law_id``/``verified`` (X11 #154) link the attachment to
+    the specific legal argument it supports and the corpus ruling, and mark whether
+    the chair verified it (the INV-AH gate the writer respects)."""
     pool = await get_pool()
     row = await pool.fetchrow(
         """
         INSERT INTO case_precedents
-            (case_id, section_id, quote, citation, chair_note, pdf_document_id, practice_area)
+            (case_id, section_id, quote, citation, chair_note, pdf_document_id,
-        VALUES ($1, $2, $3, $4, $5, $6, $7)
+             practice_area, argument_id, case_law_id, verified, verified_at)
+        VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10,
+                CASE WHEN $10 THEN now() ELSE NULL END)
         RETURNING *
         """,
-        case_id, section_id, quote, citation, chair_note, pdf_document_id, practice_area,
+        case_id, section_id, quote, citation, chair_note, pdf_document_id,
+        practice_area, argument_id, case_law_id, verified,
     )
     return dict(row)
 
 
+async def set_case_precedent_verified(
+    precedent_id: UUID, verified: bool, chair_note: str | None = None,
+) -> dict | None:
+    """Toggle the chair-verification gate on an attached precedent (#154), optionally
+    updating the chair note. Sets verified_at when verifying, clears it when un-verifying."""
+    pool = await get_pool()
+    sets = ["verified = $2", "verified_at = CASE WHEN $2 THEN now() ELSE NULL END",
+            "updated_at = now()"]
+    params: list = [precedent_id, verified]
+    if chair_note is not None:
+        sets.append(f"chair_note = ${len(params) + 1}")
+        params.append(chair_note)
+    row = await pool.fetchrow(
+        f"UPDATE case_precedents SET {', '.join(sets)} WHERE id = $1 RETURNING *",
+        *params,
+    )
+    return dict(row) if row else None
+
+
 async def list_case_precedents(case_id: UUID) -> list[dict]:
     """List all precedents attached to a case, ordered by section then creation time."""
     pool = await get_pool()
     rows = await pool.fetch(
         """
         SELECT id, case_id, section_id, quote, citation, chair_note,
-               pdf_document_id, practice_area, created_at, updated_at
+               pdf_document_id, practice_area, argument_id, case_law_id,
+               verified, verified_at, created_at, updated_at
         FROM case_precedents
         WHERE case_id = $1
         ORDER BY section_id NULLS LAST, created_at
@@ -4253,6 +4499,34 @@ async def case_number_collides(case_number: str, exclude_id: UUID) -> bool:
     ))
 
 
+# Subject-tag convention for case_law topic hubs (powers /graph topic nodes).
+# A simple multi-word Hebrew phrase is stored with underscores between words
+# ("היטל_השבחה") — the documented extraction contract (precedent_library tool
+# examples: קווי_בניין, מועד_קביעת_שומה) that ~99% of the corpus already follows.
+# The same phrase arriving with spaces ("היטל השבחה") is the same topic in
+# disguise and spawns a duplicate graph hub. Normalize at the single case_law
+# write chokepoint (G1) so no path can re-introduce the split. Tags carrying
+# punctuation/digits/dashes (e.g. "פטור מותנה — סעיף 19(ג)") are left untouched —
+# the convention covers only plain Hebrew word phrases.
+_PURE_HEBREW_WORDS = re.compile(r"^[א-ת]+(?: [א-ת]+)+$")
+
+
+def _normalize_subject_tags(tags) -> list[str]:
+    """Strip, apply the underscore convention to plain Hebrew phrases, dedup."""
+    out: list[str] = []
+    seen: set[str] = set()
+    for raw in tags or []:
+        t = str(raw).strip()
+        if not t:
+            continue
+        if _PURE_HEBREW_WORDS.match(t):
+            t = t.replace(" ", "_")
+        if t not in seen:
+            seen.add(t)
+            out.append(t)
+    return out
+
+
 async def create_external_case_law(
     case_number: str,
     case_name: str,
@@ -4270,15 +4544,27 @@ async def create_external_case_law(
     precedent_level: str = "",
     is_binding: bool = True,
     document_id: UUID | None = None,
+    citation_formatted: str = "",
 ) -> dict:
     """Insert a chair-uploaded external precedent into case_law.
 
     If a row with this ``case_number`` already exists with
     source_kind='cited_only' (auto-discovered), promote it to
     source_kind='external_upload' and fill in the missing fields.
+
+    ``citation_formatted`` seeds the מראה-מקום from the chair's typed upload-form
+    value so the field is NEVER blank between upload and metadata extraction (the
+    inline enrichment later UPGRADES it to the canonical derived form via
+    ``apply_to_record(force_citation=True)``). On a cited_only→external promotion
+    an existing non-empty value (a prior chair edit) is preserved.
     """
+    # INV-DM7: an appeals-committee source is persuasive even when uploaded via
+    # the external path — coerce to non-binding so it matches the committee
+    # invariant and case_law_committee_not_binding_check (G1).
+    if source_type == "appeals_committee":
+        is_binding = False
     pool = await get_pool()
-    tags_json = json.dumps(subject_tags or [], ensure_ascii=False)
+    tags_json = json.dumps(_normalize_subject_tags(subject_tags), ensure_ascii=False)
     async with pool.acquire() as conn:
         # Atomic upsert on the V15 partial unique index
         # uq_case_law_external_number (case_number) WHERE source_kind <> 'internal_committee'.
@@ -4295,11 +4581,12 @@ async def create_external_case_law(
                 summary, key_quote, full_text, source_url,
                 source_kind, document_id, extraction_status,
                 halacha_extraction_status, practice_area, appeal_subtype,
-                headnote, source_type, precedent_level, is_binding, content_hash
+                headnote, source_type, precedent_level, is_binding, content_hash,
+                citation_formatted
             ) VALUES (
                 $1, $2, $3, $4, $5, $6, $7, $8, $9,
                 'external_upload', $10, 'processing', 'pending',
-                $11, $12, $13, $14, $15, $16, $17
+                $11, $12, $13, $14, $15, $16, $17, $18
             )
             ON CONFLICT (case_number) WHERE source_kind <> 'internal_committee'
             DO UPDATE SET
@@ -4321,14 +4608,16 @@ async def create_external_case_law(
                 source_kind = 'external_upload',
                 extraction_status = 'processing',
                 halacha_extraction_status = 'pending',
-                content_hash = EXCLUDED.content_hash
+                content_hash = EXCLUDED.content_hash,
+                citation_formatted = COALESCE(
+                    NULLIF(case_law.citation_formatted, ''), EXCLUDED.citation_formatted)
             RETURNING *
             """,
             case_number, case_name, court, decision_date, tags_json,
             summary, key_quote, full_text, source_url,
             document_id, practice_area, appeal_subtype, headnote,
             source_type, precedent_level, is_binding,
-            _content_hash(full_text),
+            _content_hash(full_text), (citation_formatted or "").strip(),
         )
     return _row_to_case_law(row)
 
@@ -4345,7 +4634,7 @@ async def create_internal_committee_decision(
     appeal_subtype: str = "",
     subject_tags: list[str] | None = None,
     summary: str = "",
-    is_binding: bool = True,
+    is_binding: bool = False,
     document_id: UUID | None = None,
     proceeding_type: str = "ערר",
 ) -> dict:
@@ -4355,9 +4644,13 @@ async def create_internal_committee_decision(
     exist as both 'ערר' and 'בל"מ' (an extension-of-time request can be
     filed against an existing appeal with the same number).
     """
+    # INV-DM7: authority is structural for this source — an appeals-committee
+    # decision is persuasive, never binding. Coerce regardless of caller so the
+    # value can never violate case_law_committee_not_binding_check (G1).
+    is_binding = False
     pool = await get_pool()
     case_number = _canonical_case_number(case_number)
-    tags_json = json.dumps(subject_tags or [], ensure_ascii=False)
+    tags_json = json.dumps(_normalize_subject_tags(subject_tags), ensure_ascii=False)
     async with pool.acquire() as conn:
         # Atomic upsert on V15 partial unique index
         # uq_case_law_internal_number_proc (case_number, proceeding_type)
@@ -4498,6 +4791,12 @@ async def update_case_law(case_law_id: UUID, **fields) -> dict | None:
         "proceeding_type", "citation_formatted", "parties",
     }
     updates = {k: v for k, v in fields.items() if k in allowed}
+    # INV-DM7: reclassifying a row to an appeals-committee source makes it
+    # persuasive — coerce is_binding in the same patch so the UPDATE can't
+    # violate case_law_committee_not_binding_check (e.g. the Gemini metadata
+    # extractor relabels an external court_ruling as appeals_committee).
+    if updates.get("source_type") == "appeals_committee":
+        updates["is_binding"] = False
     if not updates:
         return await get_case_law(case_law_id)
 
@@ -4506,12 +4805,21 @@ async def update_case_law(case_law_id: UUID, **fields) -> dict | None:
     params: list = [case_law_id]
     for i, (k, v) in enumerate(updates.items(), start=2):
         if k == "subject_tags":
-            v = json.dumps(v or [], ensure_ascii=False)
+            v = json.dumps(_normalize_subject_tags(v), ensure_ascii=False)
         set_parts.append(f"{k} = ${i}")
         params.append(v)
     sql = f"UPDATE case_law SET {', '.join(set_parts)} WHERE id = $1 RETURNING *"
     row = await pool.fetchrow(sql, *params)
-    return _row_to_case_law(row) if row else None
+    if row is None:
+        return None
+    # `searchable` is a DERIVED completeness flag (INV-DM1). It is computed at
+    # ingest end and after the Gemini metadata extractor — but internal-committee
+    # decisions skip Gemini, so when their summary/tags/metadata are filled later
+    # (manual edit, deterministic enrichment) the flag would otherwise go stale
+    # and the row stays invisible to RAG. Recompute at this write so the derived
+    # value never drifts from the content (G1: normalize at the source).
+    await recompute_searchable(case_law_id)
+    return await get_case_law(case_law_id)
 
 
 async def set_case_law_extraction_status(case_law_id: UUID, status: str) -> None:
@@ -4920,14 +5228,24 @@ async def search_digests_semantic(
     subject_tag: str = "",
     concept_tag: str = "",
     limit: int = 10,
+    linked_only: bool | None = None,
 ) -> list[dict]:
     """Pure-semantic search over the digests radar (X12). Single vector per row
     (no chunks/halachot), so no RRF here — see X12 §6. Joins the linked ruling's
-    citation when present so the researcher sees the pointer target directly."""
+    citation when present so the researcher sees the pointer target directly.
+
+    ``linked_only``: None = all digests (default); False = only UNLINKED digests
+    (``linked_case_law_id IS NULL`` — rulings we don't hold yet, the case radar's
+    target set); True = only linked digests.
+    """
     pool = await get_pool()
     conditions = ["d.embedding IS NOT NULL"]
     params: list = [query_embedding, limit]
     idx = 3
+    if linked_only is True:
+        conditions.append("d.linked_case_law_id IS NOT NULL")
+    elif linked_only is False:
+        conditions.append("d.linked_case_law_id IS NULL")
     if practice_area:
         conditions.append(f"d.practice_area = ${idx}")
         params.append(practice_area)
@@ -6438,8 +6756,13 @@ async def refresh_verified_layer() -> dict:
     """Recompute the verified/cite_count layer from chair citations (#153).
 
     'verified' = the principle's SOURCE precedent was cited by a chair (any
-    committee decision). 'cite_count' = # distinct chair decisions citing it. This
+    committee decision) WITHOUT a negative treatment. 'cite_count' = # distinct
-    is the ONLY trust signal — never human review. Idempotent (full recompute).
+    chair decisions citing it whose treatment is NOT negative (X11 §4 / INV-COR2:
+    a precedent *distinguished*/*criticized*/*questioned*/*overruled* must never
+    gain authority from those citations). Unclassified edges (treatment='') count
+    as neutral-positive until ``classify_citation_treatments.py`` labels them, so
+    the signal degrades gracefully before classification has run. This is the ONLY
+    trust signal — never human review. Idempotent (full recompute).
     Returns {verified_principles, verified_precedents}.
     """
     pool = await get_pool()
@@ -6456,6 +6779,8 @@ async def refresh_verified_layer() -> dict:
                 "  JOIN case_law src ON src.id = pic.source_case_law_id "
                 "  WHERE src.source_kind='internal_committee' "
                 "    AND pic.cited_case_law_id IS NOT NULL "
+                "    AND coalesce(pic.treatment,'') NOT IN "
+                "        ('distinguished','criticized','questioned','overruled') "
                 "  GROUP BY pic.cited_case_law_id) "
                 "UPDATE halachot h SET verified=true, cite_count=cc.n, updated_at=now() "
                 "FROM cc WHERE h.case_law_id = cc.id")
@@ -6466,6 +6791,55 @@ async def refresh_verified_layer() -> dict:
     return {"verified_principles": row["vp"], "verified_precedents": row["vc"]}
 
 
+# X11 §4 treatment buckets (mirrors corroboration.TREATMENT_POSITIVE/NEGATIVE) —
+# kept here so the SQL layer can label a breakdown without importing the service.
+_TREATMENT_POSITIVE = ("followed", "explained")
+_TREATMENT_NEGATIVE = ("distinguished", "criticized", "questioned", "overruled")
+
+
+async def citation_authority(case_law_ids: list["UUID"]) -> dict[str, dict]:
+    """Per-precedent incoming-citation breakdown by treatment (X11 Phase 2, #154).
+
+    For each precedent id → how many DISTINCT committee decisions cite it, split
+    into positive (followed/explained), negative (distinguished/criticized/
+    questioned/overruled) and unclassified (treatment not yet labelled). This is the
+    'cited_by N (X אומצו, Y אובחנו)' authority signal surfaced to research agents so
+    they can argue authority — and avoid leaning on a precedent that was repeatedly
+    distinguished/overruled. Counts distinct sources; a source with no treatment yet
+    falls in 'unclassified'. Returns {} for ids with no incoming committee citations.
+    """
+    if not case_law_ids:
+        return {}
+    pool = await get_pool()
+    rows = await pool.fetch(
+        "SELECT pic.cited_case_law_id::text AS id, "
+        "       coalesce(NULLIF(pic.treatment, ''), 'unclassified') AS t, "
+        "       count(DISTINCT pic.source_case_law_id) AS n "
+        "FROM precedent_internal_citations pic "
+        "JOIN case_law src ON src.id = pic.source_case_law_id "
+        "WHERE src.source_kind = 'internal_committee' "
+        "  AND pic.cited_case_law_id = ANY($1::uuid[]) "
+        "GROUP BY 1, 2",
+        case_law_ids,
+    )
+    out: dict[str, dict] = {}
+    for r in rows:
+        d = out.setdefault(r["id"], {
+            "total": 0, "positive": 0, "negative": 0, "unclassified": 0,
+            "by_treatment": {},
+        })
+        t, n = r["t"], int(r["n"])
+        d["by_treatment"][t] = d["by_treatment"].get(t, 0) + n
+        d["total"] += n
+        if t in _TREATMENT_POSITIVE:
+            d["positive"] += n
+        elif t in _TREATMENT_NEGATIVE:
+            d["negative"] += n
+        else:
+            d["unclassified"] += n
+    return out
+
+
 async def list_canonical_instances(canonical_id: "UUID") -> list[dict]:
     """List all halachot (instances) sharing a canonical_id — used by the UI accordion."""
     pool = await get_pool()
@@ -8030,7 +8404,33 @@ _MP_PROVENANCE_COLS = """,
                   WHERE pic.cited_case_law_id = mp.linked_case_law_id
                     AND COALESCE(src.case_number, '') <> ''
                ) AS cited_by_precedents,
-               substring(mp.notes from 'מס''?\\s*([0-9]+)') AS yomon_number"""
+               substring(mp.notes from 'מס''?\\s*([0-9]+)') AS yomon_number,
+               -- Bridge to the corpus citation graph (G2: read-time, no stored
+               -- duplication). For an OPEN gap (no linked_case_law_id yet) find
+               -- which committee DECISIONS cite this ruling, matched on the
+               -- normalized docket number (same normalization the relinker uses:
+               -- strip to digits/dashes, slash->dash). Surfaces "צוטט ע\"י <יו\"ר>"
+               -- + the deciding case number in the UI.
+               (SELECT array_agg(DISTINCT src.chair_name ORDER BY src.chair_name)
+                  FROM precedent_internal_citations picc
+                  JOIN case_law src ON src.id = picc.source_case_law_id
+                       AND src.source_kind = 'internal_committee'
+                  WHERE split_part(mp.citation_norm, '|', 2) <> ''
+                    AND regexp_replace(replace(picc.cited_case_number, '/', '-'),
+                                       '[^0-9-]', '', 'g')
+                        = split_part(mp.citation_norm, '|', 2)
+                    AND COALESCE(src.chair_name, '') <> ''
+               ) AS cited_by_chairs,
+               (SELECT array_agg(DISTINCT src.case_number ORDER BY src.case_number)
+                  FROM precedent_internal_citations picd
+                  JOIN case_law src ON src.id = picd.source_case_law_id
+                       AND src.source_kind = 'internal_committee'
+                  WHERE split_part(mp.citation_norm, '|', 2) <> ''
+                    AND regexp_replace(replace(picd.cited_case_number, '/', '-'),
+                                       '[^0-9-]', '', 'g')
+                        = split_part(mp.citation_norm, '|', 2)
+                    AND COALESCE(src.case_number, '') <> ''
+               ) AS cited_by_decisions"""
 
 
 async def list_missing_precedents(

mcp-server/src/legal_mcp/services/digest_library.py

@@ -381,6 +381,119 @@ async def link_digest(digest_id: UUID | str, case_law_id: UUID | str) -> dict:
     }
 
 
+async def _radar_enrich(h: dict, score: float, matched_issues: list[str]) -> dict:
+    """Shape one radar hit into a chair lead: gap status + suggested action +
+    which case ISSUE(s) it answers. The action points at the underlying RULING,
+    never the digest (INV-DIG1)."""
+    cit = (h.get("underlying_citation") or "").strip()
+    gap = await db.find_missing_precedent_by_citation(cit) if cit else None
+    in_corpus = await db.find_case_law_by_citation_fuzzy(cit) if cit else None
+    if in_corpus:
+        action = "available_link"   # ruling actually IS in the corpus → just link the digest
+    elif gap and (gap.get("status") in ("uploaded", "closed")):
+        action = "fetched"          # already obtained
+    elif gap:
+        action = "gap_open"         # flagged as missing — can request a fetch
+    else:
+        action = "new_lead"         # not even flagged yet — the highest-value alert
+    return {
+        "digest_id": str(h["id"]),
+        "yomon_number": h.get("yomon_number"),
+        "headline": h.get("headline_holding") or h.get("summary") or "",
+        "underlying_citation": cit,
+        "underlying_court": h.get("underlying_court") or "",
+        "score": round(score, 3),
+        "matched_issues": matched_issues,
+        "missing_precedent_id": str(gap["id"]) if gap else None,
+        "missing_precedent_status": gap.get("status") if gap else None,
+        "action": action,
+    }
+
+
+async def case_digest_radar(
+    case_number: str,
+    limit: int = 5,
+    min_score: float = 0.45,
+) -> dict:
+    """Case-contextual digest radar (X12) — the chair-facing "שים לב" lead.
+
+    Surfaces UNLINKED digests (``linked_case_law_id IS NULL`` — rulings we don't hold
+    yet) whose topic is semantically close to THIS case, so a relevant ruling we only
+    know about via a digest doesn't fall through the cracks while the case is decided.
+
+    Query calibration: prefer the analyst's DISTILLED legal arguments (one crisp issue
+    per row — ``argument_title`` + ``legal_topic``) over the dozens of raw claims. Each
+    issue is searched separately and the leads are MERGED, so every lead is attributed
+    to the case issue(s) it answers (``matched_issues``) — far higher precision than
+    one blended query of noisy claims. Falls back to raw claims pre-aggregation
+    (``source`` reports which path ran). Each lead carries the underlying ruling's gap
+    status + a suggested action.
+
+    INV-DIG1: this is RADAR — the digest is never cited; the lead points at the
+    underlying *ruling* (fetch / upload / link), never the digest itself. Read-only.
+    """
+    case = await db.get_case_by_number(case_number)
+    if not case:
+        return {"status": "case_not_found", "case_number": case_number, "leads": [], "count": 0}
+    case_id = case["id"]
+    if isinstance(case_id, str):
+        case_id = UUID(case_id)
+    ctx = " ".join(x for x in [case.get("title") or "", case.get("appeal_subtype") or ""] if x).strip()
+
+    # Preferred source: the analyst's distilled CREAC issues (one per legal_argument).
+    issues: list[tuple[str, str]] = []  # (label, embed_text)
+    try:
+        from legal_mcp.services import argument_aggregator
+        for a in await argument_aggregator.get_legal_arguments(case_id):
+            label = (a.get("argument_title") or a.get("legal_topic") or "").strip()
+            topic = (a.get("legal_topic") or "").strip()
+            body = f"{label}. {topic}".strip(". ").strip()
+            if body:
+                issues.append((label, f"{ctx} {body}".strip()))
+    except Exception as e:  # noqa: BLE001 — arguments are optional; fall back to claims
+        logger.warning("case_digest_radar: get_legal_arguments failed for %s: %s", case_number, e)
+
+    merged: dict[str, dict] = {}  # digest_id -> {"hit", "best", "issues": set}
+    if issues:
+        source = "legal_arguments"
+        issues = issues[:25]  # already distilled — bound the per-issue search fan-out
+        vecs = await embeddings.embed_texts([t for _, t in issues], input_type="query")
+        for (label, _), vec in zip(issues, vecs):
+            for h in await db.search_digests_semantic(vec, limit=6, linked_only=False):
+                s = float(h.get("score", 0) or 0)
+                if s < min_score:
+                    continue
+                m = merged.setdefault(str(h["id"]), {"hit": h, "best": s, "issues": set()})
+                m["best"] = max(m["best"], s)
+                if label:
+                    m["issues"].add(label)
+    else:
+        # Fallback: blended query from raw claims (pre-aggregation), one search.
+        source = "claims"
+        parts = [ctx]
+        try:
+            claims = await db.get_claims(case_id)
+            parts += [(c.get("claim_text") or "") for c in claims[:20]]
+        except Exception as e:  # noqa: BLE001
+            logger.warning("case_digest_radar: get_claims failed for %s: %s", case_number, e)
+        query = " ".join(p for p in parts if p).strip()
+        if not query:
+            return {"status": "no_topic", "case_number": case_number,
+                    "source": source, "leads": [], "count": 0}
+        vec = (await embeddings.embed_texts([query], input_type="query"))[0]
+        for h in await db.search_digests_semantic(vec, limit=max(limit * 4, 20), linked_only=False):
+            s = float(h.get("score", 0) or 0)
+            if s < min_score:
+                continue
+            merged.setdefault(str(h["id"]), {"hit": h, "best": s, "issues": set()})
+
+    ordered = sorted(merged.values(), key=lambda m: -m["best"])[:max(1, limit)]
+    leads = [await _radar_enrich(m["hit"], m["best"], sorted(m["issues"])) for m in ordered]
+    return {"status": "ok", "case_number": case_number, "source": source,
+            "issues_used": [lbl for lbl, _ in issues] if source == "legal_arguments" else None,
+            "leads": leads, "count": len(leads)}
+
+
 async def relink_digest(digest_id: UUID | str) -> dict:
     """Re-run autolink for an unlinked digest. No-op if already linked / no match."""
     digest = await db.get_digest(digest_id)

mcp-server/src/legal_mcp/services/extractor.py

@@ -1,23 +1,33 @@
 """Text extraction from PDF, DOCX, DOC, and RTF files.
 
 Primary PDF extraction: PyMuPDF direct text (for born-digital PDFs).
-Fallback: Google Cloud Vision OCR (for scanned documents).
+Fallback: Mistral OCR (for scanned documents or broken OCR layers).
+
+Routing logic (document-level, not per-page):
+  1. PyMuPDF extracts text from every page.
+  2. Pages are quality-checked via _text_quality_ok().
+  3. If ALL pages pass → use PyMuPDF output (free, ~50ms, no API call).
+  4. If ANY page fails → call Mistral OCR once for the entire PDF.
+     Mistral returns per-page Markdown; page_offsets are computed from it.
+
 DOC files: converted to DOCX via LibreOffice before extraction.
-Post-processing: Hebrew abbreviation quote fixer.
+Post-processing: Hebrew abbreviation quote fixer (PyMuPDF path only;
+Mistral handles gershayim natively).
 """
 
 from __future__ import annotations
 
 import asyncio
+import base64
 import io
 import logging
 import re
 import subprocess
 import tempfile
 from pathlib import Path
-from typing import TYPE_CHECKING
 
 import fitz  # PyMuPDF
+import httpx
 from PIL import Image
 from docx import Document as DocxDocument
 from striprtf.striprtf import rtf_to_text
@@ -25,36 +35,65 @@ from striprtf.striprtf import rtf_to_text
 from legal_mcp import config
 from legal_mcp.services import storage
 
-if TYPE_CHECKING:
-    from google.cloud import vision
-
 logger = logging.getLogger(__name__)
 
-# ── Google Cloud Vision client (imported lazily — saves ~550ms at MCP startup) ──
+# ── Mistral OCR ───────────────────────────────────────────────────
 
-_vision_client: "vision.ImageAnnotatorClient | None" = None
+_MISTRAL_OCR_URL = "https://api.mistral.ai/v1/ocr"
+_MISTRAL_OCR_MODEL = "mistral-ocr-latest"
 
 
-def _get_vision_client() -> "vision.ImageAnnotatorClient":
+async def _call_mistral_ocr(path: Path) -> list[str]:
-    global _vision_client
+    """Call Mistral OCR API on a PDF. Returns per-page Markdown text list.
-    if _vision_client is None:
+
-        from google.cloud import vision
+    The Mistral response contains ``pages[i].markdown`` for each page.
-        _vision_client = vision.ImageAnnotatorClient(
+    If the response has fewer pages than the PDF, trailing pages are
-            client_options={"api_key": config.GOOGLE_CLOUD_VISION_API_KEY}
+    padded with empty strings by the caller.
+    """
+    if not config.MISTRAL_API_KEY:
+        raise RuntimeError(
+            "MISTRAL_API_KEY not configured — cannot OCR scanned PDF. "
+            "Set the env var in Coolify."
         )
-    return _vision_client
+
+    pdf_b64 = base64.b64encode(path.read_bytes()).decode()
+
+    async with httpx.AsyncClient(timeout=300.0) as client:
+        resp = await client.post(
+            _MISTRAL_OCR_URL,
+            headers={
+                "Authorization": f"Bearer {config.MISTRAL_API_KEY}",
+                "Content-Type": "application/json",
+            },
+            json={
+                "model": _MISTRAL_OCR_MODEL,
+                "document": {
+                    "type": "document_url",
+                    "document_url": f"data:application/pdf;base64,{pdf_b64}",
+                },
+                "include_image_base64": False,
+            },
+        )
+
+    if resp.status_code != 200:
+        raise RuntimeError(
+            f"Mistral OCR returned {resp.status_code}: {resp.text[:400]}"
+        )
+
+    pages = resp.json().get("pages", [])
+    return [p.get("markdown", "") for p in pages]
 
 
 # ── Hebrew text quality detection ────────────────────────────────
 
-_HEBREW_RE = re.compile(r'[\u0590-\u05FF]')
+_HEBREW_RE = re.compile(r'[֐-׿]')
 _WORD_RE = re.compile(r'\S+')
 
 
 def _text_quality_ok(text: str) -> bool:
-    """Check if extracted text is real content vs broken OCR layer.
+    """Check if PyMuPDF-extracted text is genuine Hebrew legal content.
 
-    Returns True if text appears to be genuine Hebrew legal content.
+    Returns True if text appears to be real content.
     Broken OCR layers from scanned PDFs often have:
     - Very short words / single-character fragments
     - Each word on its own line (high words-per-line ratio)
@@ -64,26 +103,19 @@ def _text_quality_ok(text: str) -> bool:
     if len(words) < 10:
         return False
 
-    # Average word length — real Hebrew words avg 4-6 chars.
     avg_len = sum(len(w) for w in words) / len(words)
     if avg_len < 2.5:
         return False
 
-    # Percentage of single-character "words"
     single_char_pct = sum(1 for w in words if len(w) == 1) / len(words)
     if single_char_pct > 0.4:
         return False
 
-    # Words per line — broken OCR puts each word on its own line.
-    # Real text has 5-15 words per line; broken OCR has ~1-2.
     lines = [l for l in text.split("\n") if l.strip()]
-    if lines:
+    if lines and len(words) / len(lines) < 3.0:
-        words_per_line = len(words) / len(lines)
+        return False
-        if words_per_line < 3.0:
-            return False
 
-    # Hebrew character ratio among letter characters
+    letters = re.findall(r'[a-zA-Z֐-׿]', text)
-    letters = re.findall(r'[a-zA-Z\u0590-\u05FF]', text)
     if letters:
         hebrew_pct = sum(1 for c in letters if _HEBREW_RE.match(c)) / len(letters)
         if hebrew_pct < 0.5:
@@ -92,7 +124,7 @@ def _text_quality_ok(text: str) -> bool:
     return True
 
 
-# ── Hebrew abbreviation quote fixer ──────────────────────────────
+# ── Hebrew abbreviation quote fixer (PyMuPDF path only) ──────────
 
 _HEBREW_ABBREV_FIXES: dict[str, str] = {
     'עוהייד': 'עוה"ד',
@@ -111,50 +143,133 @@ _HEBREW_ABBREV_FIXES: dict[str, str] = {
     'יחייד': 'יח"ד',
     'בייכ': 'ב"כ',
     # Patterns where double-yod (יי) substitutes for gershayim (״) in born-digital PDFs
-    'בליימ': 'בל"מ',   # בקשה להארכת מועד — appears in RTL legal docs
+    'בליימ': 'בל"מ',
-    'תמייא': 'תמ"א',   # תכנית מתאר ארצית
+    'תמייא': 'תמ"א',
 }
 
 _ABBREV_PATTERN = re.compile(
     '|'.join(re.escape(k) for k in sorted(_HEBREW_ABBREV_FIXES, key=len, reverse=True))
 )
 
-# Matches Hebrew law year abbreviations where gershayim was encoded as double-yod.
-# e.g. תשכייה → תשכ"ה,  תשנייב → תשנ"ב
 _HEBREW_YEAR_RE = re.compile(r'(תש[א-ת]+)יי([א-ת])')
 
 
 def _fix_hebrew_quotes(text: str) -> str:
-    """Fix known Hebrew abbreviation quote replacements.
+    """Fix gershayim encoded as double-yod in born-digital PDFs."""
-
-    Applied to both Google Vision OCR output and direct PyMuPDF extraction —
-    some born-digital PDFs encode gershayim (״) as double-yod (יי), producing
-    the same corruption patterns as OCR.
-    """
     text = _ABBREV_PATTERN.sub(lambda m: _HEBREW_ABBREV_FIXES[m.group()], text)
     text = _HEBREW_YEAR_RE.sub(r'\1"\2', text)
     return text
 
 
-# ── Extraction ───────────────────────────────────────────────────
+# ── Page joining ──────────────────────────────────────────────────
-
 
 # Separator used when joining per-page text. Constant so chunker /
 # retrofit can reproduce the join when computing page offsets.
 PAGE_SEPARATOR = "\n\n"
 
 
+def _join_pages(pages_text: list[str]) -> tuple[str, list[int]]:
+    """Join per-page text with PAGE_SEPARATOR while recording start offsets."""
+    offsets: list[int] = []
+    parts: list[str] = []
+    cursor = 0
+    for i, pg in enumerate(pages_text):
+        offsets.append(cursor)
+        parts.append(pg)
+        cursor += len(pg)
+        if i < len(pages_text) - 1:
+            parts.append(PAGE_SEPARATOR)
+            cursor += len(PAGE_SEPARATOR)
+    return "".join(parts), offsets
+
+
+# ── PDF extraction ────────────────────────────────────────────────
+
+
+async def _extract_pdf(path: Path) -> tuple[str, int, list[int]]:
+    """Extract text from PDF using document-level routing.
+
+    Stage 1 — PyMuPDF pre-screen (free, ~50ms, no API call):
+        Run on every page. Collect per-page text; flag pages where
+        PyMuPDF returns < 50 chars or _text_quality_ok() fails
+        (scanned, blank, or broken embedded OCR layer).
+
+    Stage 2 — Mistral OCR (triggered when any page fails Stage 1):
+        Send the entire PDF to Mistral once. Use its per-page Markdown
+        for ALL pages — consistent source, no mixed plain/Markdown formats.
+        Mistral handles gershayim natively; no quote-fix applied.
+
+    Page offsets are always computed so the chunker can attribute each
+    chunk to its source page number (multimodal hybrid retrieval).
+    """
+    doc = fitz.open(str(path))
+    page_count = len(doc)
+
+    # Stage 1: PyMuPDF pre-screen
+    pymupdf_pages: list[str] = []
+    failed: list[int] = []
+
+    for i in range(page_count):
+        text = doc[i].get_text().strip()
+        if len(text) > 50 and _text_quality_ok(text):
+            pymupdf_pages.append(_fix_hebrew_quotes(text))
+        else:
+            pymupdf_pages.append("")
+            failed.append(i)
+
+    doc.close()
+
+    if not failed:
+        logger.debug(
+            "PDF %s: all %d pages digital — PyMuPDF only", path.name, page_count
+        )
+        joined, offsets = _join_pages(pymupdf_pages)
+        return joined, page_count, offsets
+
+    # Stage 2: Mistral OCR for entire document
+    logger.info(
+        "PDF %s: %d/%d pages failed quality check → Mistral OCR",
+        path.name, len(failed), page_count,
+    )
+    mistral_pages = await _call_mistral_ocr(path)
+
+    # Pad if Mistral returns fewer pages than PyMuPDF counted
+    while len(mistral_pages) < page_count:
+        mistral_pages.append("")
+
+    joined, offsets = _join_pages(mistral_pages[:page_count])
+    return joined, page_count, offsets
+
+
+def page_at_offset(offset: int, page_offsets: list[int]) -> int:
+    """Return the 1-based page number containing a given char offset.
+
+    page_offsets[i] is the start of page (i+1) in the joined text.
+    """
+    if not page_offsets:
+        return 1
+    page = 1
+    for i, start in enumerate(page_offsets):
+        if start <= offset:
+            page = i + 1
+        else:
+            break
+    return page
+
+
+# ── Public entry point ────────────────────────────────────────────
+
+
 async def extract_text(file_path: str) -> tuple[str, int, list[int] | None]:
     """Extract text from a document file.
 
     Returns:
         ``(text, page_count, page_offsets)`` where:
-        - ``text``: concatenated extracted text
+        - ``text``: extracted text. Plain text for PyMuPDF path;
+          Markdown for Mistral path (tables, ``##`` headers preserved).
         - ``page_count``: number of pages (0 for non-PDF)
-        - ``page_offsets``: ``page_offsets[i]`` = char start offset of
+        - ``page_offsets``: char start of each page inside ``text``,
-          page (i+1) inside ``text``. ``None`` for non-PDFs (where the
+          or ``None`` for non-PDF formats
-          notion of pages doesn't apply). Used by the chunker to assign
-          a ``page_number`` to each chunk.
     """
     path = Path(file_path)
     suffix = path.suffix.lower()
@@ -173,95 +288,11 @@ async def extract_text(file_path: str) -> tuple[str, int, list[int] | None]:
         raise ValueError(f"Unsupported file type: {suffix}")
 
 
-def _join_pages(pages_text: list[str]) -> tuple[str, list[int]]:
+# ── Non-PDF formats ───────────────────────────────────────────────
-    """Join per-page text with PAGE_SEPARATOR while recording the start
-    offset of each page in the joined output."""
-    offsets: list[int] = []
-    parts: list[str] = []
-    cursor = 0
-    for i, pg in enumerate(pages_text):
-        offsets.append(cursor)
-        parts.append(pg)
-        cursor += len(pg)
-        if i < len(pages_text) - 1:
-            parts.append(PAGE_SEPARATOR)
-            cursor += len(PAGE_SEPARATOR)
-    return "".join(parts), offsets
-
-
-async def _extract_pdf(path: Path) -> tuple[str, int, list[int]]:
-    """Extract text from PDF.
-
-    Try direct text first, fall back to Google Cloud Vision for scanned
-    or broken-OCR pages.
-    """
-    doc = fitz.open(str(path))
-    page_count = len(doc)
-    pages_text: list[str] = []
-
-    for page_num in range(page_count):
-        page = doc[page_num]
-        text = page.get_text().strip()
-
-        if len(text) > 50 and _text_quality_ok(text):
-            pages_text.append(_fix_hebrew_quotes(text))
-            logger.debug("Page %d: direct extraction (%d chars, quality OK)", page_num + 1, len(text))
-        else:
-            reason = "insufficient text" if len(text) <= 50 else "low quality OCR layer"
-            logger.info("Page %d: Google Vision OCR (%s)", page_num + 1, reason)
-            pix = page.get_pixmap(dpi=300)
-            img_bytes = pix.tobytes("png")
-            ocr_text = await asyncio.to_thread(
-                _ocr_with_google_vision, img_bytes, page_num + 1
-            )
-            pages_text.append(ocr_text)
-
-    doc.close()
-    joined, offsets = _join_pages(pages_text)
-    return joined, page_count, offsets
-
-
-def page_at_offset(offset: int, page_offsets: list[int]) -> int:
-    """Look up the page number containing a given char offset.
-
-    page_offsets[i] is the start of page (i+1) in the joined text;
-    a chunk starting at ``offset`` belongs to the highest-indexed page
-    whose start is ``<= offset``. Returns 1-based page number.
-    """
-    if not page_offsets:
-        return 1
-    # Linear scan is fine — page_offsets is short (≤ ~200 for our PDFs).
-    page = 1
-    for i, start in enumerate(page_offsets):
-        if start <= offset:
-            page = i + 1
-        else:
-            break
-    return page
-
-
-def _ocr_with_google_vision(image_bytes: bytes, page_num: int) -> str:
-    """OCR a single page image using Google Cloud Vision API."""
-    from google.cloud import vision  # lazy: keeps MCP startup fast
-    client = _get_vision_client()
-    image = vision.Image(content=image_bytes)
-
-    response = client.document_text_detection(
-        image=image,
-        image_context=vision.ImageContext(language_hints=["he"]),
-    )
-
-    if response.error.message:
-        raise RuntimeError(
-            f"Google Vision error on page {page_num}: {response.error.message}"
-        )
-
-    text = response.full_text_annotation.text if response.full_text_annotation else ""
-    return _fix_hebrew_quotes(text)
 
 
 def _extract_doc(path: Path) -> str:
-    """Extract text from legacy .doc file by converting to .docx via LibreOffice."""
+    """Extract text from legacy .doc via LibreOffice → DOCX conversion."""
     with tempfile.TemporaryDirectory() as tmp_dir:
         # Isolate the LibreOffice user profile per call: headless soffice
         # locks a single shared profile, so concurrent .doc conversions would
@@ -296,13 +327,13 @@ def _extract_rtf(path: Path) -> str:
 
 
 # ── Multimodal page rendering (V9) ───────────────────────────────
+# Unchanged — multimodal embedding always uses PyMuPDF-rendered images
+# regardless of whether text extraction used PyMuPDF or Mistral.
 
 
 def _pixmap_to_pil(pix: fitz.Pixmap) -> Image.Image:
-    """Convert a PyMuPDF pixmap to PIL.Image (RGB) without going through
+    """Convert a PyMuPDF pixmap to PIL.Image (RGB)."""
-    PNG bytes. Faster than tobytes('png') → Image.open()."""
     if pix.alpha:
-        # Drop alpha channel — voyage multimodal expects RGB.
         pix = fitz.Pixmap(pix, 0)
     return Image.frombytes("RGB", (pix.width, pix.height), pix.samples)
 
@@ -314,12 +345,9 @@ def render_pages_for_multimodal(
     thumbnail_dir: Path | None = None,
 ) -> list[tuple[Image.Image, Path | None]]:
     """Render each PDF page as PIL.Image at ``embed_dpi`` for the
-    multimodal embedder, and optionally save a smaller JPEG thumbnail
+    multimodal embedder, and optionally save JPEG thumbnails.
-    at ``thumb_dpi`` to ``thumbnail_dir`` for UI preview.
 
     Returns ``[(pil_image, thumb_path_or_None), ...]`` in page order.
-    The full-DPI image stays in memory only — only the thumbnail is
-    persisted to disk.
     """
     src = Path(pdf_path)
     if not src.is_file():
@@ -338,17 +366,12 @@ def render_pages_for_multimodal(
             thumb_path: Path | None = None
             if thumbnail_dir is not None and thumb_dpi:
                 thumb_path = thumbnail_dir / f"p{page_num:03d}.jpg"
-                # Downsample the same render rather than re-rendering
-                # with PyMuPDF — far faster.
                 ratio = thumb_dpi / embed_dpi
                 thumb_size = (
                     max(1, int(img.width * ratio)),
                     max(1, int(img.height * ratio)),
                 )
                 thumb = img.resize(thumb_size, Image.Resampling.LANCZOS)
-                # Persist the thumbnail (a DERIVED, regenerable artifact)
-                # through the storage layer (INV-STG1). Under the filesystem
-                # backend it lands at thumb_path exactly as before.
                 _tbuf = io.BytesIO()
                 thumb.save(_tbuf, "JPEG", quality=75, optimize=True)
                 try:
@@ -366,44 +389,28 @@ def render_pages_for_multimodal(
     return out
 
 
-# ── Nevo preamble stripping ──────────────────────────────────────
+# ── Nevo preamble stripping ───────────────────────────────────────
 
 _NEVO_MARKERS = ("ספרות:", "חקיקה שאוזכרה:", "מיני-רציו:", "פסקי דין שאוזכרו:",
                  "כתבי עת:", "הועתק מנבו")
 
-# Markers for where the actual decision body begins (everything before is Nevo
-# preamble: bibliography + מיני-רציו). Two families:
-#   - ועדת ערר / district openings (בפנינו / הערר שבנדון / ...)
-#   - COURT-RULING openings (#86.1): a פסק-דין header or the authoring judge's
-#     line. Without these, Nevo court judgments — exactly the ones carrying a
-#     מיני-רציו — slipped through unstripped (e.g. בג"ץ 1764/05).
-#
-# #86.2 hardening — two over-strip bugs found while backfilling:
-#   1. ``פסק-דין`` headers are often markdown-wrapped (``**פסק  דין**``); the old
-#      ``^פסק[- ]דין`` required the keyword to be the very first char of the line
-#      and allowed only one separator, so it missed the header and fell through
-#      to a citation 32K deep (עמ"נ 50567-07-21). We now tolerate leading
-#      markdown/whitespace and 0-3 separators.
-#   2. Bare ``השופט``/``הנשיא`` matched *citations* ("השופט מ' חשין, פסקה 23"),
-#      stripping real decision body. The authoring-judge line ends with a COLON
-#      ("השופט י' עמית:"); citations use a comma. We now require the colon.
 _DECISION_START = re.compile(
     r"^[ \t>*_#]{0,6}(?:"
     r"בפנינו|לפנינו|לפניי|הערר שבנדון|ועדת הערר לתכנון|רקע עובדתי|עסקינן|"
-    r"פסק[ \t\-]{0,3}די(?:ן|נו)|"  # פסק-דין / פסק דין / **פסק  דין** header (final-nun ן vs דינו)
+    r"פסק[ \t\-]{0,3}די(?:ן|נו)|"
-    r"(?:כב(?:וד)?['׳\"]?\s*)?(?:ה?שופט[ת]?|ה?נשיא[ה]?|המשנה לנשיא)\s+[^\n,]{1,40}:"  # author line → colon
+    r"(?:כב(?:וד)?['׳\"]?\s*)?(?:ה?שופט[ת]?|ה?נשיא[ה]?|המשנה לנשיא)\s+[^\n,]{1,40}:"
     r")",
     re.MULTILINE,
 )
 
 
 def strip_nevo_preamble(text: str) -> str:
-    """Remove Nevo database preamble (bibliography, legislation, mini-ratio) from decision text.
+    """Remove Nevo database preamble (bibliography, legislation, mini-ratio).
 
     Returns the original text unchanged if no preamble is detected.
+    Works on both plain text (PyMuPDF) and Markdown (Mistral) since
+    _DECISION_START already tolerates leading ``[ \t>*_#]{0,6}``.
     """
-    # Window wide enough to catch the Nevo markers even when a long court/parties
-    # header precedes them (court rulings push חקיקה שאוזכרה:/מיני-רציו: down).
     head = text[:1500]
     if not any(marker in head for marker in _NEVO_MARKERS):
         return text
@@ -419,17 +426,7 @@ _RATIO_MARKER = "מיני-רציו:"
 
 
 def extract_nevo_ratio(text: str) -> str:
-    """Return the Nevo מיני-רציו block (editorial holdings summary), or ''.
+    """Return the Nevo מיני-רציו block (editorial holdings summary), or ''."""
-
-    The mini-ratio is Nevo's own headnote — a concise, professionally-written
-    list of the holdings. We capture it *before* :func:`strip_nevo_preamble`
-    discards it, to serve as a free gold-set for benchmarking how well our
-    halacha extractor covers the real holdings (#86.3).
-
-    The block runs from the ``מיני-רציו:`` marker to whichever comes first:
-    the decision body (``_DECISION_START``) or the next preamble marker
-    (bibliography / legislation). Returns '' when there is no mini-ratio.
-    """
     if not text:
         return ""
     start = text.find(_RATIO_MARKER)
@@ -437,9 +434,6 @@ def extract_nevo_ratio(text: str) -> str:
         return ""
     body = text[start + len(_RATIO_MARKER):]
 
-    # End at the earliest of: decision body start, or a following preamble
-    # marker (ספרות: / חקיקה שאוזכרה: / ...). Both are measured relative to
-    # the ratio body so we never run past it into the judgment itself.
     end = len(body)
     dm = _DECISION_START.search(body)
     if dm:

mcp-server/src/legal_mcp/services/ingest.py

@@ -233,6 +233,17 @@ async def ingest_document(
             await db.request_metadata_extraction(case_law_id)
             await db.request_halacha_extraction(case_law_id)
             await db.recompute_searchable(case_law_id)
+            # Citations resolve to a corpus row only at extraction time. A ruling
+            # uploaded AFTER it was first cited leaves orphan edges (NULL link),
+            # so re-link them to this freshly-ingested row now — the citation
+            # graph self-heals instead of permanently under-counting (G1). Non-fatal.
+            try:
+                from legal_mcp.services import citation_extractor as _ce
+                relinked = await _ce.relink_orphan_citations(case_law_id)
+                if relinked:
+                    logger.info("relinked %d orphan citation(s) -> %s", relinked, case_law_id)
+            except Exception as e:  # noqa: BLE001 — relink is best-effort
+                logger.warning("citation relink failed (non-fatal): %s", e)
 
             await progress("completed", 100,
                            f"נקלט: {stored_chunks} chunks. חילוץ הלכות ומטא-דאטה ממתינים בתור.")

mcp-server/src/legal_mcp/services/internal_decisions.py

@@ -28,14 +28,14 @@ logger = logging.getLogger(__name__)
 INTERNAL_DECISIONS_DIR = Path(config.DATA_DIR) / "internal-decisions"
 
 _VALID_PRACTICE_AREAS = frozenset({"", "rishuy_uvniya", "betterment_levy", "compensation_197"})
-_VALID_DISTRICTS = frozenset({"", "ירושלים", "מרכז", "תל אביב", "צפון", "דרום", "ארצי"})
+_VALID_DISTRICTS = frozenset({"", "ירושלים", "מרכז", "תל אביב", "חיפה", "צפון", "דרום", "ארצי"})
 
 _COURT_TO_DISTRICT = [
     ("ירושלים", "ירושלים"),
     ("תל אביב", "תל אביב"),
     ('ת"א', "תל אביב"),
     ("מרכז", "מרכז"),
-    ("חיפה", "צפון"),
+    ("חיפה", "חיפה"),
     ("צפון", "צפון"),
     ("דרום", "דרום"),
     ("ארצי", "ארצי"),
@@ -77,7 +77,7 @@ async def _create_internal_record(**kw) -> dict:
         appeal_subtype=(kw.get("appeal_subtype") or "").strip(),
         subject_tags=list(kw.get("subject_tags") or []),
         summary=(kw.get("summary") or "").strip(),
-        is_binding=kw.get("is_binding", True),
+        is_binding=kw.get("is_binding", False),  # INV-DM7: committee = persuasive
         document_id=kw.get("document_id"),
         proceeding_type=kw.get("proceeding_type") or "ערר",
     )
@@ -108,7 +108,7 @@ async def ingest_internal_decision(
     appeal_subtype: str = "",
     subject_tags: list[str] | None = None,
     summary: str = "",
-    is_binding: bool = True,
+    is_binding: bool = False,  # INV-DM7: committee sources are persuasive
     file_path: str | Path | None = None,
     text: str | None = None,
     document_id: UUID | None = None,
@@ -229,6 +229,7 @@ async def migrate_from_external_corpus(dry_run: bool = False) -> dict:
             await conn.execute(
                 """UPDATE case_law
                    SET source_kind = 'internal_committee',
+                       is_binding = FALSE,  -- INV-DM7: committee = persuasive
                        district = CASE WHEN $2 <> '' THEN $2 ELSE district END
                    WHERE id = $1""",
                 row["id"], district,

mcp-server/src/legal_mcp/services/precedent_library.py

@@ -68,7 +68,11 @@ def _external_staging_subdir(inputs: dict) -> str:
 
 
 async def _create_external_record(**kw) -> dict:
-    """Adapter: maps canonical inputs (citation) to create_external_case_law(case_number)."""
+    """Adapter: maps canonical inputs (citation) to create_external_case_law(case_number).
+
+    The chair's typed citation seeds ``citation_formatted`` so the מראה-מקום is never
+    blank before metadata extraction; the inline enrichment upgrades it to the
+    canonical form (see ``ingest_precedent``)."""
     return await db.create_external_case_law(
         case_number=kw["citation"].strip(),
         case_name=kw["case_name"],
@@ -84,6 +88,7 @@ async def _create_external_record(**kw) -> dict:
         precedent_level=kw.get("precedent_level", ""),
         is_binding=kw.get("is_binding", True),
         document_id=kw.get("document_id"),
+        citation_formatted=kw["citation"].strip(),
     )
 
 
@@ -126,10 +131,28 @@ async def ingest_precedent(
         "appeal_subtype": appeal_subtype, "subject_tags": subject_tags,
         "is_binding": is_binding, "headnote": headnote, "summary": summary,
     }
-    return await ingest.ingest_document(
+    result = await ingest.ingest_document(
         _EXTERNAL_SPEC, inputs=inputs, file_path=file_path,
         document_id=document_id, progress=progress,
     )
+    # Inline metadata enrichment (Gemini, in-container — gemini_session is direct
+    # REST, no local CLI). UPGRADES the seeded provisional citation_formatted to the
+    # canonical derived form (parties + reporter + date) so the מראה-מקום is correct
+    # the moment the chair opens /precedents/[id] — not deferred to the local drainer.
+    # force_citation=True overwrites the seed only; chair edits aren't reachable yet
+    # (row just created). Best-effort: on no key / API failure the seed remains and
+    # the queued metadata drain (request_metadata_extraction, already set) is the
+    # fallback. Halacha extraction is NOT touched here — it stays local (claude_session).
+    cid = result.get("case_law_id") if isinstance(result, dict) else None
+    if cid:
+        try:
+            from legal_mcp.services import precedent_metadata_extractor as _pme
+            r = await _pme.extract_and_apply(UUID(str(cid)), force_citation=True)
+            if r.get("status") == "completed":
+                await db.set_case_law_metadata_status(UUID(str(cid)), "completed")
+        except Exception as e:  # noqa: BLE001 — enrichment is best-effort; drainer is fallback
+            logger.warning("inline metadata enrichment failed for %s (drainer will retry): %s", cid, e)
+    return result
 
 
 async def reextract_halachot(
@@ -360,7 +383,10 @@ async def reextract_metadata(
     appeal_subtype, and case_name when it equals the citation). User
     values are preserved.
 
-    **MCP-tool-only path** — same constraint as :func:`reextract_halachot`.
+    **Container-safe** — unlike :func:`reextract_halachot` (claude CLI, host-only),
+    metadata extraction runs on Gemini Flash over REST (GOOGLE_GEMINI_API_KEY), so
+    this path is callable from the FastAPI container too. The final-decision
+    enrollment loop (``_enroll_final_in_library``) calls it inline on upload.
     """
     from legal_mcp.services import precedent_metadata_extractor
 
@@ -411,6 +437,37 @@ async def get_precedent(case_law_id: UUID | str) -> dict | None:
         return None
     record["halachot"] = await db.list_halachot(case_law_id=case_law_id, limit=500)
     record["related_cases"] = await db.get_case_law_relations(case_law_id)
+    # Auto-detected citation graph: which decisions cite THIS ruling (incoming
+    # edges in precedent_internal_citations). Reuses the canonical query (G2) —
+    # no parallel resolution path. Shaped to match the front-end RelatedCase row
+    # so the existing "ציטוטים מקושרים" panel renders them with no visual change.
+    from legal_mcp.services import citation_extractor as _ce
+    raw = await _ce.list_citations_to_case_law(case_law_id)
+    record["incoming_citations"] = [
+        {
+            "id": r["source_case_law_id"],
+            "case_number": r.get("source_case_number") or r.get("cited_case_number") or "",
+            "case_name": r.get("source_case_name") or "",
+            "court": r.get("source_court") or "",
+            "precedent_level": r.get("source_precedent_level") or "",
+            "chair_name": r.get("source_chair_name") or "",
+            "date": (
+                r["source_date"].isoformat()
+                if r.get("source_date") is not None else None
+            ),
+            "treatment": r.get("treatment") or "",
+            "confidence": r.get("confidence"),
+        }
+        for r in raw
+    ]
+    # Authority signal (X11 Phase 2, #154): how the citing committee decisions
+    # TREATED this ruling (followed/distinguished/…) — surfaced so the chair (and
+    # research agents via the tool output) can argue authority and avoid leaning on
+    # a repeatedly-distinguished precedent.
+    record["cited_by"] = (await db.citation_authority([case_law_id])).get(
+        str(case_law_id),
+        {"total": 0, "positive": 0, "negative": 0, "unclassified": 0, "by_treatment": {}},
+    )
     return record
 
 

mcp-server/src/legal_mcp/services/precedent_metadata_extractor.py

@@ -260,6 +260,7 @@ async def apply_to_record(
     case_law_id: UUID | str,
     suggested: dict,
     overwrite_case_number: bool = False,
+    force_citation: bool = False,
 ) -> dict:
     """Merge suggested metadata into the case_law row, filling ONLY empty fields.
 
@@ -274,6 +275,13 @@ async def apply_to_record(
 
     overwrite_case_number: when True, update case_number from case_number_clean
       even if the field already has a value (used for one-time migration enrichment).
+
+    force_citation: when True, (re)assemble citation_formatted even if the field
+      is non-empty — used by the at-upload inline enrichment to UPGRADE the seeded
+      provisional citation (the raw chair input) to the canonical derived form. The
+      write still happens only when the deterministic assembly SUCCEEDS (a missing
+      component → no write → the seed is preserved). The drainer keeps the default
+      (False) so a chair's manual edit in /precedents/[id] is never clobbered.
     """
     if isinstance(case_law_id, str):
         case_law_id = UUID(case_law_id)
@@ -482,7 +490,7 @@ async def apply_to_record(
     # source_type/district/proceeding_type/parties). Only fill when empty so chair
     # edits in /precedents/[id] are preserved; abstains (no write) when a component
     # is missing.
-    if not (record.get("citation_formatted") or "").strip():
+    if force_citation or not (record.get("citation_formatted") or "").strip():
         eff = {**record, **fields_to_update}
         eff_parties = (
             fields_to_update.get("parties") or record.get("parties") or ""
@@ -505,6 +513,7 @@ async def apply_to_record(
 async def extract_and_apply(
     case_law_id: UUID | str,
     overwrite_case_number: bool = False,
+    force_citation: bool = False,
 ) -> dict:
     """Convenience wrapper: extract → merge into row → return summary."""
     suggested = await extract_metadata(case_law_id)
@@ -523,7 +532,11 @@ async def extract_and_apply(
             "status": "extraction_failed" if has_text else "no_metadata",
             "fields": [],
         }
-    result = await apply_to_record(case_law_id, suggested, overwrite_case_number=overwrite_case_number)
+    result = await apply_to_record(
+        case_law_id, suggested,
+        overwrite_case_number=overwrite_case_number,
+        force_citation=force_citation,
+    )
     if result["updated"]:
         await db.recompute_searchable(case_law_id)
     return {

mcp-server/src/legal_mcp/services/style_exemplars.py

@@ -0,0 +1,83 @@
+"""Block-level style-exemplar extraction (channel B of Style Acquisition).
+
+Splits one of Dafna's decisions into section→paragraph units, embeds them
+(Voyage), and stores them in `style_exemplars` so the writer can retrieve real
+block-level prose by section/outcome/practice_area (07-learning §0.2 channel B).
+
+This is the SINGLE source of truth for exemplar extraction (G2): both the
+one-time backfill (`scripts/backfill_style_exemplars.py`) and the live
+final-enrollment path (`_enroll_final_in_library`) call `extract_and_store`.
+Before this, exemplars were frozen at the seed backfill — new finals enrolled
+into style_corpus but were never broken into exemplars, so the richest style
+channel never grew. Embedding is Voyage-over-REST → container-safe.
+"""
+from __future__ import annotations
+
+import logging
+
+from legal_mcp.services import db, embeddings
+from legal_mcp.services.chunker import _split_into_sections
+
+logger = logging.getLogger(__name__)
+
+# chunker section_type → style_exemplars.section
+_SECTION_MAP = {
+    "facts": "background",
+    "appellant_claims": "claims",
+    "respondent_claims": "claims",
+    "legal_analysis": "discussion",
+    "conclusion": "summary",
+    "ruling": "summary",
+    "intro": "other",
+    "other": "other",
+}
+
+MIN_WORDS = 25      # skip tiny fragments
+MAX_WORDS = 450     # skip over-long blobs (likely un-split)
+MAX_PER_SECTION = 15
+
+
+def _paragraphs(section_text: str) -> list[str]:
+    """Split a section into paragraph units (blank-line separated; fall back to lines)."""
+    raw = [p.strip() for p in section_text.split("\n\n")]
+    if len(raw) <= 1:
+        raw = [p.strip() for p in section_text.split("\n")]
+    out = []
+    for p in raw:
+        wc = len(p.split())
+        if MIN_WORDS <= wc <= MAX_WORDS:
+            out.append(p)
+    return out[:MAX_PER_SECTION]
+
+
+def units_for(full_text: str) -> list[tuple[str, str]]:
+    """(section, paragraph) units for a decision — pure, no I/O."""
+    units: list[tuple[str, str]] = []
+    for section_type, section_text in _split_into_sections(full_text or ""):
+        section = _SECTION_MAP.get(section_type, "other")
+        for para in _paragraphs(section_text):
+            units.append((section, para))
+    return units
+
+
+async def extract_and_store(
+    decision_number: str, source: str, full_text: str,
+    practice_area: str = "", outcome: str = "",
+) -> int:
+    """Idempotently (re)build a decision's block-level style exemplars: split →
+    embed → replace. Returns the number of exemplars stored. Raises on failure;
+    callers in best-effort paths (enroll) should wrap in try/except."""
+    units = units_for(full_text)
+    if not units:
+        return 0
+    texts = [u[1] for u in units]
+    vecs = await embeddings.embed_texts(texts, input_type="document")
+    await db.delete_style_exemplars(decision_number, source)
+    for (section, para), vec in zip(units, vecs):
+        await db.insert_style_exemplar(
+            decision_number=decision_number, source=source,
+            practice_area=practice_area, outcome=outcome,
+            section=section, paragraph_text=para, word_count=len(para.split()),
+            embedding=vec,
+        )
+    return len(units)

mcp-server/src/legal_mcp/tools/digests.py

@@ -170,3 +170,30 @@ async def digest_process_pending(limit: int = 20) -> str:
     except Exception as e:
         return _err(str(e))
     return _ok(result)
+
+
+async def digest_radar(case_number: str, limit: int = 5, min_score: float = 0.45) -> str:
+    """רדאר-יומונים הקשרי-לתיק (X12) — "שים לב" ליו"ר.
+
+    מחזיר יומונים **לא-מקושרים** (פס"ד שעוד אין לנו בקורפוס) שהנושא שלהם קרוב
+    סמנטית לתיק הזה — כדי שפס"ד רלוונטי שמוכר רק דרך יומון לא ייפול בין הכיסאות
+    בזמן הכרעת התיק. נושא-התיק נבנה מ-title + appeal_subtype + טענות-הצדדים, ומותאם
+    מול יומונים לא-מקושרים. כל ליד נושא את מראה-המקום של הפס"ד המקורי, סטטוס-הפער
+    (new_lead / gap_open / fetched / available_link) ופעולה מוצעת.
+
+    INV-DIG1: זהו radar — היומון לעולם אינו מצוטט; הליד מצביע על **הפס"ד** (להזמין
+    משיכה / להעלות / לקשר), לא על היומון. read-only.
+
+    Args:
+        case_number: מספר התיק (למשל "8124-09-24").
+        limit: מספר לידים מקסימלי.
+        min_score: סף-דמיון סמנטי (0-1) לסינון רעש.
+    """
+    if not case_number.strip():
+        return _err("case_number חובה")
+    try:
+        result = await digest_library.case_digest_radar(
+            case_number.strip(), limit=max(1, int(limit)), min_score=float(min_score))
+    except Exception as e:
+        return _err(str(e))
+    return _ok(result)

mcp-server/src/legal_mcp/tools/precedent_library.py

@@ -300,6 +300,20 @@ async def search_precedent_library(
         limit=limit,
         include_halachot=include_halachot,
     )
+    # X11 Phase 2 (#154): attach the incoming-citation authority breakdown so the
+    # research agent can WEIGH and ARGUE authority ("הלכה שאומצה ב-N החלטות ועדת-ערר")
+    # — and steer clear of a precedent that committees repeatedly distinguished /
+    # overruled. Batched: one query for the whole result page.
+    try:
+        ids = {str(r.get("case_law_id")) for r in results if r.get("case_law_id")}
+        if ids:
+            auth = await db.citation_authority([UUID(i) for i in ids])
+            for r in results:
+                cb = auth.get(str(r.get("case_law_id")))
+                if cb:
+                    r["cited_by"] = cb
+    except Exception:  # noqa: BLE001 — authority is an additive signal; never break search
+        pass
     elapsed_ms = int((time.perf_counter() - t0) * 1000)
     telemetry.log_search_bg(
         search_type="precedent_library",

mcp-server/src/legal_mcp/tools/workflow.py

@@ -394,13 +394,107 @@ async def record_chair_feedback(
         lesson_extracted=lesson_extracted,
     )
 
+    # Auto-flow chair-authored STYLE feedback to the writer (closes the dead
+    # chair_feedback→lesson chain — 27 feedback rows had produced 0 lessons). The
+    # chair is the highest authority, so a style correction she writes flows
+    # immediately — the chair IS the gate. It rides the SAME discussion_rules channel
+    # promote uses (db.append_global_rule, G2), reaching every block, without the
+    # style_corpus coupling decision_lessons require. SUBSTANCE feedback
+    # (missing_content/factual_error/other) is case-specific → recorded only.
+    # (INV-LRN1 graduated gate; 07-learning §1.2.)
+    _STYLE_FB = {"style", "wrong_tone", "wrong_structure"}
+    flowed = 0
+    if lesson_extracted.strip() and category in _STYLE_FB:
+        try:
+            flowed = await db.append_global_rule(
+                "discussion_rules", "universal", [lesson_extracted.strip()],
+            )
+        except Exception as e:
+            logger.warning("chair-feedback auto-flow failed for %s: %s", case_number, e)
+
+    msg = f"הערה נרשמה בהצלחה. קטגוריה: {category}."
+    if flowed:
+        msg += " הלקח (סגנון) זרם אוטומטית לכותב."
     return ok({
         "feedback_id": str(feedback_id),
+        "flowed_to_writer": bool(flowed),
         "next_steps": [
             "כדי להפיק לקח מההערה, הפעל: analyze_chair_feedback",
             "כדי לסמן כמטופל: resolve_chair_feedback",
         ],
-    }, message=f"הערה נרשמה בהצלחה. קטגוריה: {category}.")
+    }, message=msg)
+
+
+_CURATOR_FINDING_CATEGORIES = {"style", "structure", "lexicon", "tabular", "general"}
+# Agent tags ([סגנון]/[מבנה]/[לקסיקון משפטי]/[טבלאי]) → decision_lessons.category
+_CURATOR_TAG_TO_CATEGORY = {
+    "סגנון": "style", "מבנה": "structure",
+    "לקסיקון משפטי": "lexicon", "לקסיקון": "lexicon", "טבלאי": "tabular",
+}
+
+
+async def record_curator_findings(case_number: str, findings: list[dict]) -> str:
+    """לכידת ממצאי-האוצֵר כ-decision_lessons מובְנים (source='curator', proposed) — INV-LRN3.
+
+    האוצֵר מזהה דפוסי-סגנון בקריאת הסופי; עד כה הם חיו רק כהערה ארעית בערוץ-התגובות
+    (אובד, לא נסקר). כאן הם נתפסים מבנית כך שיופיעו בטאב ״אוצֵר״ ויעברו שער-יו"ר (INV-LRN1/G10).
+    האוצֵר נשאר read-only על התוכן — הרישום הוא הצעה הממתינה לאישור, לא שינוי-קול.
+
+    Args:
+        case_number: מספר התיק הסופי (= decision_number בקורפוס-הסגנון).
+        findings: רשימת ממצאים, כל אחד {"text": "...", "category"/"tag": "..."}.
+                  category ∈ style/structure/lexicon/tabular/general (או tag עברי).
+    """
+    if not findings:
+        return err("findings ריק — אין ממצאים לרשום.")
+    corpus_id = await db.get_style_corpus_id_by_decision(case_number)
+    if not corpus_id:
+        return err(
+            f"לא נמצאה רשומת style_corpus ל-{case_number} — ודא שהסופי נקלט לקורפוס-הסגנון "
+            "(enroll_style_corpus) לפני רישום ממצאים."
+        )
+    # Dedup against lessons already on this corpus (any source) — re-running §A
+    # must not pile duplicates (INV-LRN3 reliability).
+    existing = {(_norm(r["lesson_text"])) for r in await db.list_decision_lessons(corpus_id)}
+    written, skipped_dup, skipped_empty = [], 0, 0
+    for f in findings:
+        text = (f.get("text") or "").strip()
+        if not text:
+            skipped_empty += 1
+            continue
+        if _norm(text) in existing:
+            skipped_dup += 1
+            continue
+        raw_cat = (f.get("category") or f.get("tag") or "general").strip()
+        category = _CURATOR_TAG_TO_CATEGORY.get(raw_cat, raw_cat)
+        if category not in _CURATOR_FINDING_CATEGORIES:
+            category = "general"
+        row = await db.add_decision_lesson(
+            corpus_id,
+            lesson_text=text,
+            category=category,
+            source="curator",
+            created_by="curator",
+            review_status="proposed",
+        )
+        if row:
+            written.append(str(row["id"]))
+            existing.add(_norm(text))
+    return ok({
+        "corpus_id": str(corpus_id),
+        "written": len(written),
+        "skipped_duplicate": skipped_dup,
+        "skipped_empty": skipped_empty,
+        "lesson_ids": written,
+    }, message=(
+        f"נרשמו {len(written)} ממצאי-אוצֵר (source=curator, ממתינים לשער-יו\"ר ב-/training). "
+        f"{skipped_dup} כפילויות דולגו."
+    ))
+
+
+def _norm(s: str) -> str:
+    """Normalize lesson text for dedup — collapse whitespace, strip."""
+    return " ".join((s or "").split())
 
 
 async def list_chair_feedback(

mcp-server/tests/test_adapter_profiles.py

@@ -0,0 +1,21 @@
+"""Adapter-profile compatibility gates for Paperclip migration."""
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+
+_SCRIPT = Path(__file__).resolve().parents[2] / "scripts" / "adapter_profiles.py"
+_spec = importlib.util.spec_from_file_location("adapter_profiles", _SCRIPT)
+profiles = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(profiles)
+
+
+def test_codex_local_profile_accepts_openai_model_ids():
+    assert profiles.model_matches_provider("gpt-5.3-codex", "codex_local")
+    assert profiles.model_matches_provider("o4-mini", "codex_local")
+    assert profiles.model_matches_provider("codex-mini-latest", "codex_local")
+
+
+def test_codex_local_profile_rejects_foreign_model_ids():
+    assert not profiles.model_matches_provider("claude-opus-4-8", "codex_local")
+    assert not profiles.model_matches_provider("gemini-3.1-pro-preview", "codex_local")

scripts/SCRIPTS.md

@@ -15,6 +15,7 @@
 | `pc.sh` | bash | **wrapper לכל קריאות Paperclip API מסוכנים** — מוסיף Authorization, X-Paperclip-Run-Id (audit trail), Content-Type, base URL. תחביר: `pc.sh <METHOD> <PATH> [BODY_JSON]`. אסור `curl` ישיר ל-`$PAPERCLIP_API_URL`. ראה `HEARTBEAT.md §0`. counterpart ב-Python: `web/paperclip_api.py`. | נקרא ע"י סוכנים |
 | `sync_agents_across_companies.py` | python | **סנכרון סוכנים מ-CMP (1xxx, master) ל-CMPA (8xxx, mirror)** — Gap #25. משווה adapter_config (model/timeout/instructions/skills/etc), runtime_config (heartbeat), ושדות top-level (budget/metadata/icon/title/role). מסנן אוטומטית local skills שלא קיימים ב-mirror. לוגיקת subset (mirror יכול להחזיק יותר skills כי ה-API מוסיף required runtime skills). תומך `--verify`/`--dry-run`/`--apply [--only NAME]`. גיבוי אוטומטי. דורש `PAPERCLIP_BOARD_API_KEY`. **להריץ אחרי כל שינוי הגדרות ב-CMP.** **⚠ אם `adapter_type` שונה בין CMP ל-CMPA — `--apply` מדלג על הסוכן; `--verify` מדווח אותו רם כ-DRIFT.** בעת מעבר adapter (למשל ל-`deepseek_local`) חובה לעדכן ידנית בשתי החברות. **`--verify` יוצא exit≠0 על כל drift** (needs-sync / adapter-mismatch / missing-in-mirror) — שמיש כ-gate ל-cron/CI (GAP-21/FU-8a). | ידני אחרי כל שינוי |
 | `fix_paperclipai_skills_drift.py` | python | סקריפט חד-פעמי (בוצע 2026-05-04) שניקה drift על `paperclipai/*` skills בין CMP ל-CMPA. הסיר `paperclip-dev` מכל 14 הסוכנים, ודאג ש-`paperclip-converting-plans-to-tasks` קיים רק על CEO ו-analyst. תומך `--apply` (ברירת מחדל: dry-run). דורש `PAPERCLIP_BOARD_API_KEY`. נשמר לרפרנס למקרה שhdrift חוזר. | חד-פעמי (בוצע) |
+| `classify_citation_treatments.py` | python | **סיווג-טיפול לקצוות-ציטוט (X11 Phase 2, #154)** — לכל קצה ב-`precedent_internal_citations` (החלטת-ועדה מצטטת תקדים) מסווג את ה-`treatment` מתוך ה-`match_context` דרך `corroboration.classify_treatment` (Opus 4.8 @ xhigh, claude_session **מקומי** — לא בקונטיינר): followed/explained=חיובי, distinguished/criticized/questioned/overruled=שלילי. ממלא `precedent_internal_citations.treatment` כך ש-`refresh_verified_layer` לא יספור ציטוט שלילי כסמכות (INV-COR2) ו-`db.citation_authority` יציג פירוק לסוכנים. אידמפוטנטי (מדלג על מסווגים). `--apply`/`--limit N`/`--case-law-id UUID`. **אחרי `--apply` הרץ `build_verified_layer.py`.** דורש `HOME=/home/chaim`. | ידני / אחרי גלי-ציטוט חדשים |
 | `adapter_profiles.py` | python (module) | **רישום-פרופילי-אדפטר** — מקור-אמת יחיד ל-3 צירי-הכשל של מעבר-אדפטר: provider/default_model, instructions_mode (`file_path` בטוח-frontmatter מול `content_arg` ששובר `---`), ו-tool_config (`gemini_global` excludeTools / `frontmatter` / `hermes` / `codex_home`). כולל `codex_local` עם משפחת מודלי OpenAI/Codex (`gpt-*`, `o3*`, `o4*`, `codex-*`). מיובא ע"י `migrate_agent_adapter.py`. הוספת אדפטר עתידי = רשומה אחת. לא מורץ ישירות. | תשתית |
 | `migrate_agent_adapter.py` | python | **מעבר-אדפטר בטוח לכל סוכן ← כל אדפטר, בשתי החברות יחד (INV-MC1)**. מיישב model↔provider, גורס frontmatter לעותק `.generated/<name>.nofm.md` ל-content_arg adapters (אחרת קריסת `gemini --prompt`/`hermes -q` על `---`), ומשחרר excludeTools גלובלי של gemini (`--relax-tools`). `--check` (preflight בלבד, exit≠0 על שגיאה — שער FU-8a) / `--apply` / `--revert` (שחזור מדויק מ-sidecar `data/adapter-migration-state.json`) / `--verify` (מסמן מצב לא-תואם/א-סימטרי, exit≠0). `--agent "<שם>"\|all --to <adapter> [--model X] [--relax-tools]`. PATCH דרך `/api/agents/{id}` (לא DB). דורש `PAPERCLIP_BOARD_API_KEY`. הרץ עם `mcp-server/.venv/bin/python`. **fallback-חירום כשנגמרים טוקני-Claude; החזר ל-claude_local כשחוזרים.** | ידני לפי צורך |
 
@@ -152,6 +153,7 @@
 
 | Script | Type | Purpose | Scheduled |
 |--------|------|---------|-----------|
+| `ocr_benchmark_mistral.py` | python | **בנצ'מרק OCR — Mistral מול Google Vision** (מחקר חד-פעמי, הוביל למעבר ל-Mistral). מוריד מסמכים מ-MinIO, קורא טקסט קיים מה-DB, שולח ל-Mistral OCR, מחשב מטריקות (כיסוי/ניקיון/עברית%), שומר דוח ל-`data/audit/ocr-benchmark-mistral.md` + טקסטים גולמיים ל-`data/audit/ocr-benchmark-raw/`. הרצה: `mcp-server/.venv/bin/python scripts/ocr_benchmark_mistral.py`. דורש `MISTRAL_API_KEY` ו-mcli alias `legalminio`. **ממצא:** Mistral מנצח ב-4/5 תיקים בדוגמה ומטפל נכון ב-OCR שבור (1044-03-26 שהחזיר "English garbage" ב-Vision). | חד-פעמי (בוצע 2026-06-27) |
 | `backfill_missing_precedents.py` | python | **הזנת `missing_precedents` פתוחים לתור-האחזור (X13)** — מסווג כל פער-פתוח; עליון-סדרתי→Tier-0(supremedecisions), נט-format→Tier-1; ועדת-ערר/לא-מזוהה→דילוג. יוצר `court_fetch_jobs` (idempotent). `--apply` (ברירת-מחדל dry-run). אחרי הרצה: drain-court-fetch קולט. | ידני (חד-פעמי/לפי-צורך) |
 | `derive_missing_from_cited_only.py` | python | **#143 — איחוד cited_only↔missing_precedents (G2)**: גוזר רשומת `missing_precedents` 'open' לכל stub `cited_only` (פסיקה מצוטטת ללא טקסט), כך ש-31 ה-stubs מופיעים בדף "פסיקה חסרה" (היו היו חפיפה≈0). (1) backfill `citation_norm` (מפתח-dedup designator-aware — `court_citation.citation_dedup_key`) ל-291 הקיימים; (2) לכל stub → `create_missing_precedent(discovery_source='cited_only', linked_case_law_id=stub, notes=מצטטים)` עם dedup. `linked_case_law_id`=זהות-קנונית-ידועה, `status='open'` עד העלאת-טקסט (→ promote-in-place דרך ON CONFLICT). אידמפוטנטי, dry-run / `--apply`. הרצה: `HOME=/home/chaim mcp-server/.venv/bin/python scripts/derive_missing_from_cited_only.py --apply`. | חד-פעמי / re-runnable |
 | `backfill_digest_missing_precedents.py` | python | **#136 — חיבור יומונים-לא-מקושרים ל"פסיקה חסרה"**: לכל digest עם `underlying_citation` ו-`linked_case_law_id IS NULL` (461) מריץ את `digest_library.try_autolink` הקנוני (G2) — מקשר אם אפשר, אחרת פותח gap: ערר/בל"מ/unknown → `missing_precedent` (discovery_source='digest', dedup designator-aware), פס"ד בתי-משפט → `court_fetch_job` (X13). dry-run מציג פילוח-tier (369 ערר + 21 unknown → MP; 71 fetchable → court_fetch). אידמפוטנטי. הרצה: `HOME=/home/chaim mcp-server/.venv/bin/python scripts/backfill_digest_missing_precedents.py --apply`. | חד-פעמי / re-runnable |
@@ -198,7 +200,7 @@
 | `drain_digests.py` | python | ריקון תור ההעשרה של יומונים (X12): מעבד כל digest בסטטוס `pending` דרך `digest_library.enrich_digest` (חילוץ-LLM Sonnet + embedding + autolink). מקבילי (CONCURRENCY=3, env-tunable), idempotent. מוסיף `~/.local/bin` ל-PATH כדי שה-claude CLI יימצא תחת cron. בודק דגל `drain_controls('legal-digest-drain')` ב-startup → no-op כשכבוי מ-/operations. | דרך `legal-digest-drain.config.cjs` (pm2 cron) + ידני אחרי backfill. חלופת-MCP: `digest_process_pending` |
 | `legal-digest-drain.config.cjs` | pm2/js | **תזמון כל שעתיים של `drain_digests.py`** (cron `12 */2 * * *`, `DIGEST_DRAIN_CRON` לעקיפה; דקת-הצתה `:12` כדי לא לחלוק דקה עם metadata-drain `:00` — מונע deadlock של DDL-המיגרציה) — הועבר מ-crontab של המערכת ל-pm2 כדי שיופיע ויהיה שליט בדף `/operations` (הרץ-עכשיו/הפעל/כבה). `autorestart:false` (one-shot per tick). דורש claude CLI + `VOYAGE_API_KEY`. התקנה: `pm2 start scripts/legal-digest-drain.config.cjs && pm2 save`. | pm2 cron (host-side) |
 
-| `renumber_cases.py` | python | **מיגרציה חד-פעמית (בוצעה 2026-06-12)** — תיקון 11 מספרי-תיקים לפורמט קנוני `NNNN-MM-YY` (הוספת ספרות-חודש; 1046-26→1024-02-26 תיקון-סידורי). רץ על ה-host (לא בקונטיינר): DB pool של האפליקציה + `mcli` (MinIO) + Gitea API + Paperclip DB. אטומי per-case עם גיבוי ל-`data/audit/` ואימות-אחרי. FK-ים על `cases.id` (UUID) לא נגעו; משכתב כל עמודה עם `cases/{old}/` (file_path **וגם** image_thumbnail_path שהוא storage-key בלי `/data`), מנרמל זהות חוצת-קורפוס (case_law/style_corpus/style_exemplars/citations — לא תוכן/full_text), מעביר מפתחי-MinIO ב-3 buckets (legal-immutable=WORM copy-only), משנה-שם repo ב-Gitea, ומעדכן שם-פרויקט ב-Paperclip. dry-run כברירת-מחדל; `--apply --tier clean\|archive`. **מיצוי — לא להריץ שוב** (ה-MAPPING היסטורי). | חד-פעמי — בוצע |
+| `renumber_cases.py` | python | **מיגרציה חד-פעמית (בוצעה 2026-06-12)** — תיקון 11 מספרי-תיקים לפורמט קנוני `NNNN-MM-YY` (הוספת ספרות-חודש; 1046-26→1024-02-26 תיקון-סידורי). רץ על ה-host (לא בקונטיינר): DB pool של האפליקציה + `mcli` (MinIO) + Gitea API + Paperclip DB. אטומי per-case עם גיבוי ל-`data/audit/` ואימות-אחרי. FK-ים על `cases.id` (UUID) לא נגעו; משכתב כל עמודה עם `cases/{old}/` (file_path **וגם** image_thumbnail_path שהוא storage-key בלי `/data`), מנרמל זהות חוצת-קורפוס (case_law/style_corpus/style_exemplars/citations — לא תוכן/full_text), מעביר מפתחי-MinIO ב-3 buckets (legal-immutable=WORM copy-only), משנה-שם repo ב-Gitea, ומעדכן ב-Paperclip את **שם-הפרויקט + `plugin_state.legal-case-number` + כותרות-issues** (3 המשטחים ש-`get_case_issues` נשען עליהם — בלי (b)+(c) ה-issues נשארים על המספר הישן ו-run-learning/run-halacha מדלגים בשקט עם "no_issue"; תוקן 2026-06-28 אחרי שהפער התגלה ב-8137-11-24). dry-run כברירת-מחדל; `--apply --tier clean\|archive`. **מיצוי — לא להריץ שוב** (ה-MAPPING היסטורי); לתבנית-עתידית: ה-Paperclip-step המעודכן הוא הרפרנס. | חד-פעמי — בוצע |
 
 ## סקריפטים שנמחקו (git history בלבד)
 

scripts/backfill_style_exemplars.py

@@ -19,40 +19,15 @@ import argparse
 import asyncio
 import logging
 
-from legal_mcp.services import db, embeddings
+from legal_mcp.services import db
-from legal_mcp.services.chunker import _split_into_sections
+from legal_mcp.services import style_exemplars as sx
 
 logging.basicConfig(level=logging.INFO, format="%(message)s")
 log = logging.getLogger("backfill_exemplars")
 
-# chunker section_type → style_exemplars.section
+# Section mapping + paragraph splitting now live in the shared service
-_SECTION_MAP = {
+# (legal_mcp.services.style_exemplars) so the backfill and the live
-    "facts": "background",
+# final-enrollment path use ONE extraction implementation (G2).
-    "appellant_claims": "claims",
-    "respondent_claims": "claims",
-    "legal_analysis": "discussion",
-    "conclusion": "summary",
-    "ruling": "summary",
-    "intro": "other",
-    "other": "other",
-}
-
-MIN_WORDS = 25      # skip tiny fragments
-MAX_WORDS = 450     # skip over-long blobs (likely un-split)
-MAX_PER_SECTION = 15
-
-
-def _paragraphs(section_text: str) -> list[str]:
-    """Split a section into paragraph units (blank-line separated; fall back to lines)."""
-    raw = [p.strip() for p in section_text.split("\n\n")]
-    if len(raw) <= 1:
-        raw = [p.strip() for p in section_text.split("\n")]
-    out = []
-    for p in raw:
-        wc = len(p.split())
-        if MIN_WORDS <= wc <= MAX_WORDS:
-            out.append(p)
-    return out[:MAX_PER_SECTION]
 
 
 async def _gather_sources() -> list[dict]:
@@ -94,28 +69,18 @@ async def main(apply: bool) -> None:
 
     total_paras = 0
     for src in sources:
-        units: list[tuple[str, str]] = []  # (section, paragraph)
+        n = len(sx.units_for(src["full_text"]))
-        for section_type, section_text in _split_into_sections(src["full_text"]):
+        if not n:
-            section = _SECTION_MAP.get(section_type, "other")
-            for para in _paragraphs(section_text):
-                units.append((section, para))
-        if not units:
             continue
-        total_paras += len(units)
+        total_paras += n
-        log.info("  %-14s %-16s → %d פסקאות", src["source"], src["decision_number"], len(units))
+        log.info("  %-14s %-16s → %d פסקאות", src["source"], src["decision_number"], n)
         if not apply:
             continue
-
+        await sx.extract_and_store(
-        await db.delete_style_exemplars(src["decision_number"], src["source"])
+            decision_number=src["decision_number"], source=src["source"],
-        texts = [u[1] for u in units]
+            full_text=src["full_text"], practice_area=src["practice_area"],
-        vecs = await embeddings.embed_texts(texts, input_type="document")
+            outcome=src["outcome"],
-        for (section, para), vec in zip(units, vecs):
+        )
-            await db.insert_style_exemplar(
-                decision_number=src["decision_number"], source=src["source"],
-                practice_area=src["practice_area"], outcome=src["outcome"],
-                section=section, paragraph_text=para, word_count=len(para.split()),
-                embedding=vec,
-            )
 
     if apply:
         cov = await db.count_style_exemplars()

scripts/classify_citation_treatments.py

@@ -0,0 +1,99 @@
+"""Classify the TREATMENT of each internal citation edge (X11 Phase 2, #154).
+
+Each row in ``precedent_internal_citations`` records that a committee decision
+cited a precedent, with the surrounding ``match_context``. Until now the edge's
+``treatment`` column was empty, so the verified/authority layer counted every
+citation as if it were positive — a precedent *distinguished* N times got the
+same authority boost as one *followed* N times (an INV-COR2 violation).
+
+This script fills ``treatment`` per edge by classifying the context with
+``corroboration.classify_treatment`` (Opus 4.8 @ xhigh via the local
+claude_session bridge — LOCAL ONLY, the claude CLI is not in the container):
+
+    followed | explained        → positive (counts toward authority)
+    distinguished | criticized |
+    questioned | overruled       → negative (never counts; overruled = demote)
+
+Scope: only LINKED edges (``cited_case_law_id IS NOT NULL``) with an empty
+``treatment`` and a non-empty ``match_context``. Idempotent — a second run skips
+rows already classified. After applying, run ``scripts/build_verified_layer.py``
+(or ``db.refresh_verified_layer``) so the treatment-aware count takes effect.
+
+Run (dry-run, default — classifies and PRINTS, writes nothing):
+    HOME=/home/chaim mcp-server/.venv/bin/python scripts/classify_citation_treatments.py
+Apply:
+    HOME=/home/chaim mcp-server/.venv/bin/python scripts/classify_citation_treatments.py --apply
+Options:
+    --limit N            classify at most N edges (smoke test)
+    --case-law-id UUID   restrict to citations TO this one precedent
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import os
+import sys
+from uuid import UUID
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "mcp-server", "src"))
+
+from legal_mcp.services import corroboration, db  # noqa: E402
+
+
+async def _pending(limit: int | None, case_law_id: str | None) -> list[dict]:
+    pool = await db.get_pool()
+    where = ["cited_case_law_id IS NOT NULL", "coalesce(treatment,'') = ''",
+             "coalesce(match_context,'') <> ''"]
+    params: list = []
+    if case_law_id:
+        params.append(UUID(case_law_id))
+        where.append(f"cited_case_law_id = ${len(params)}")
+    sql = (f"SELECT id, cited_case_number, match_context "
+           f"FROM precedent_internal_citations WHERE {' AND '.join(where)} "
+           f"ORDER BY created_at")
+    if limit:
+        sql += f" LIMIT {int(limit)}"
+    rows = await pool.fetch(sql, *params)
+    return [dict(r) for r in rows]
+
+
+async def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--apply", action="store_true", help="write changes (default: dry-run)")
+    ap.add_argument("--limit", type=int, default=None)
+    ap.add_argument("--case-law-id", type=str, default=None)
+    args = ap.parse_args()
+
+    rows = await _pending(args.limit, args.case_law_id)
+    print(f"קצוות לא-מסווגים לעיבוד: {len(rows)}\n")
+
+    pool = await db.get_pool()
+    counts: dict[str, int] = {}
+    errors = 0
+    for r in rows:
+        try:
+            t = await corroboration.classify_treatment(
+                r["cited_case_number"] or "", r["match_context"] or "")
+        except Exception as e:  # noqa: BLE001 — one bad row must not abort the batch
+            errors += 1
+            print(f"  ✗ [error] {r['cited_case_number']}: {type(e).__name__}: {e}")
+            continue
+        counts[t] = counts.get(t, 0) + 1
+        sign = "＋" if corroboration.is_positive(t) else ("－" if corroboration.is_negative(t) else "·")
+        print(f"  {sign} {t:<14} {r['cited_case_number']}")
+        if args.apply:
+            await pool.execute(
+                "UPDATE precedent_internal_citations SET treatment = $2 WHERE id = $1",
+                r["id"], t,
+            )
+
+    print(f"\nסיכום טיפול: {counts}  שגיאות={errors}"
+          + ("" if args.apply else "   (dry-run — לא נכתב)"))
+    if args.apply:
+        print("הרץ עכשיו: scripts/build_verified_layer.py (או db.refresh_verified_layer) "
+              "כדי שהספירה מודעת-הטיפול תיכנס לתוקף.")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

scripts/exec_summary_1043.py

@@ -0,0 +1,145 @@
+"""Ad-hoc: executive summary (סיכום מנהלים) DOCX for case 1043-02-26.
+Reuses the dafna decision template styles (David font + RTL) via the
+analysis_docx_exporter helpers. One-off prep document for chaim's meeting
+with the chair — NOT a decision draft.
+"""
+from pathlib import Path
+import sys
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "mcp-server" / "src"))
+
+from docx import Document
+from legal_mcp.services.analysis_docx_exporter import (
+    TEMPLATE_PATH,
+    _clear_body,
+    _add_paragraph,
+    _add_runs_with_inline_bold,
+    _mark_paragraph_rtl,
+    _mark_run_rtl,
+)
+
+CASE = "1043-02-26"
+OUT = Path(f"/home/chaim/legal-ai/data/cases/{CASE}/exports/סיכום-מנהלים-v1.docx")
+
+
+def H1(doc, t):
+    _add_paragraph(doc, t, "Heading 1")
+
+
+def H2(doc, t):
+    _add_paragraph(doc, t, "Heading 2")
+
+
+def P(doc, t):
+    p = doc.add_paragraph(style="Normal")
+    _add_runs_with_inline_bold(p, t)
+    _mark_paragraph_rtl(p)
+    return p
+
+
+def BULLET(doc, t):
+    p = doc.add_paragraph(style="List Paragraph")
+    _add_runs_with_inline_bold(p, t)
+    _mark_paragraph_rtl(p)
+    return p
+
+
+def LABEL(doc, label, value):
+    p = doc.add_paragraph(style="Normal")
+    r = p.add_run(label + ": ")
+    r.bold = True
+    _mark_run_rtl(r)
+    r2 = p.add_run(value)
+    _mark_run_rtl(r2)
+    _mark_paragraph_rtl(p)
+    return p
+
+
+def main():
+    doc = Document(str(TEMPLATE_PATH))
+    _clear_body(doc)
+
+    H1(doc, "סיכום מנהלים — הכנה לדיון")
+    P(doc, "**ערר 1043-02-26 — הקמת מתקן למיון פסולת (אתר \"קומפוסט דלילה\", תכנית מי/1030)**")
+    P(doc, "מסמך הכנה פנימי לקראת דיון עם יו\"ר הוועדה. אינו החלטה ואינו טיוטת החלטה.")
+
+    H2(doc, "פרטי התיק")
+    LABEL(doc, "סוג הערר", "רישוי ובנייה (1xxx) — ערר על סירוב בקשה להיתר")
+    LABEL(doc, "עוררים (מבקשי ההיתר)", "קיבוץ נחשון; חברת אלקטרה אקו גרין פארק")
+    LABEL(doc, "משיבים", "הוועדה המקומית מטה יהודה; מושבי גפן/תירוש/כפר הריף; קיבוצי כפר מנחם/רבדים; מועצה אזורית יואב (מתנגדים)")
+    LABEL(doc, "מושא הערר", "החלטת הוועדה המקומית מטה יהודה מיום 9.2.26 לסרב לבקשה (מס' 20240972)")
+    LABEL(doc, "המקרקעין", "גוש 5093 ח\"ל 4 מגרש 1, אתר \"דלילה\" (~137 דונם), מצפון לכפר מנחם, ליד כביש 383")
+    LABEL(doc, "תקן ביקורת", "ועדת הערר כמוסד תכנון בעל סמכות מקורית — שיקול דעת תכנוני עצמאי; ביקורת רחבה יותר בחלק המשפטי-פרשני")
+
+    H2(doc, "מהות המחלוקת")
+    P(doc, "מבקשי ההיתר מבקשים להקים מתקן מיון פסולת/קומפוסטציה הכולל מבנה קומפוסטציה אחוד (~19,118 מ\"ר עיקרי), משטחי הבשלה פתוחים ובריכה תפעולית, בצירוף שתי הקלות: הגבהת גובה מ-8 ל-20 מ', והגדלת תכסית מ-3.5% ל-22%. הוועדה המקומית סירבה, בקובעה כי השינוי מהותי ומקומו בעדכון התכנית ולא בהליך רישוי. השאלה המרכזית: האם הבקשה תואמת את תכנית מי/1030, או חורגת ממנה באופן המחייב תיקון תכנית.")
+
+    H2(doc, "החלטת הוועדה המקומית (מושא הערר)")
+    P(doc, "**דחתה** את טענת המתנגדים לפקיעת התכנית (בהסתמך על ע\"א 3213/97 נקר) ואת טענת השימושים האסורים.")
+    P(doc, "**קיבלה** והפכה לבסיס הסירוב: (1) היקף הבינוי חורג משלב א' לפי הוראת השלביות; (2) הבינוי המונוליטי שונה דרמטית מנספח הבינוי; (3) משטחי ההבשלה הפתוחים מנוגדים לסעיף 6.9 (טיפול במבנים סגורים) — שהפרתו מוגדרת בתכנית כסטייה ניכרת.")
+
+    H2(doc, "טענות סף (לדיון תמציתי)")
+    BULLET(doc, "**פקיעת תכנית** — סעיף הפקיעה (5 שנים) הושמט מהנוסח המאושר (ראו ממצאי ניתוח-העומק).")
+    BULLET(doc, "**מיצוי התנגדות / \"מועד ב'\"** — האם טענות שהוכרעו בשלב התכנית מועלות מחדש בשלב הרישוי.")
+    BULLET(doc, "**זכות עמידה** — מועד חתימת חוזה החכירה מול מועד פתיחת הבקשה (הוכרע: קיבוץ נחשון חוכר רשום).")
+    BULLET(doc, "**פגמי פרסום / היעדר יידוע** (ס' 149(א)(2א)) — היעדר תשריט, אי-יידוע ועדה מקומית יואב.")
+    BULLET(doc, "**איחור בהגשת ההתנגדויות** — מול אינטרס ההסתמכות (רכישת קרקע ב-35 מיליון ₪).")
+
+    H2(doc, "חמש הסוגיות המהותיות")
+
+    P(doc, "**סוגיה 1 (מכריעה) — שלביות הביצוע: בינוי מול היקף פעילות**")
+    BULLET(doc, "השאלה: האם הוראת השלביות (ס' 7.1) חלה על הבינוי הפיזי או רק על היקף ההפעלה?")
+    BULLET(doc, "עמדות: עוררים — נוגעת להיקף הפעילות; ועדה/משיבים — חלה על הבינוי (4 מבנים ≈ 8,000 מ\"ר מול ~19,118 מבוקשים).")
+    BULLET(doc, "לאן נוטה: **לטובת הוועדה** — לשון \"הקמת 4 מבני קומפוסט\" מתייחסת לבינוי; נספחים 06, 11 מצביעים על מנגנון פיילוט מדורג מהותי.")
+    BULLET(doc, "תקדים: עע\"מ 10089/07 אירוס הגלבוע (אין לעקוף תכנון נדרש דרך היתר).")
+
+    P(doc, "**סוגיה 2 (מכריעה) — משטחי ההבשלה וסעיף 6.9: \"טיפול\" מול \"אחסון\"**")
+    BULLET(doc, "השאלה: האם ההבשלה הפתוחה היא \"טיפול\" החוסה תחת 6.9 (→ סטייה ניכרת החוסמת הקלה, ס' 151), או \"אחסון\"?")
+    BULLET(doc, "עמדות: עוררים (ד\"ר ענבר, נספח 22) — אחסון מוצר סופי נטול ריח; ועדה — חלק בלתי נפרד מהטיפול.")
+    BULLET(doc, "לאן נוטה: **נחלשה לוועדה** — נספח 07 §122-127: ועדת המשנה לעררים אישרה הבשלה פתוחה כ\"נכון וסביר\" (אך כינתה זאת \"שלב אחרון של טיפול\"). מתח פרשני אמיתי.")
+    BULLET(doc, "תקדים: עע\"מ 402/03 עמותת העצמאים (ס' 151 — סטייה ניכרת).")
+
+    P(doc, "**סוגיה 3 — ההקלות בגובה ובתכסית + טענת \"טעות סופר\"**")
+    BULLET(doc, "השאלה: האם ההקלות בגדרי הקלה או סטייה ניכרת? והאם התכסית 3.5% היא \"טעות סופר\"?")
+    BULLET(doc, "לאן נוטה: טענת טעות הסופר **התחזקה מאוד** — נספח 12 (מופקד) מראה ש-3.5% נגזרה מתפיסה שהחממות \"אינן שטח לבניה\"; נספחים 06 §14 ו-07 §130 הורו לתקן; הסכם רמ\"י נוקב ב-30,280 מ\"ר.")
+    BULLET(doc, "תקדים: עמ\"נ 25955-11-22 ברק-רחביה (פרשנות הרמונית); בג\"ץ 2667/17 מטה בנימין (\"פרשנות אפשרית\").")
+
+    P(doc, "**סוגיה 4 — תצורת הבינוי מול נספח בינוי מנחה ונספח נופי מחייב**")
+    BULLET(doc, "השאלה: האם המבנה המונוליטי + הבריכה חורגים ממרחב הגמישות של נספח מנחה, לנוכח הנספח הנופי המחייב?")
+    BULLET(doc, "תקדים: **ערר 1033-25 אבו גוש** (תקדים דפנה ישיר — נספח בינוי מנחה אינו המלצה בלבד); בג\"ץ 6525/15 עמק שווה.")
+
+    P(doc, "**סוגיה 5 — שיקולים זרים / NIMBY בהחלטת הסירוב**")
+    BULLET(doc, "השאלה: האם הסירוב נגוע בשיקולים זרים? (עשוי להתייתר אם הבחינה העצמאית מכריעה).")
+    BULLET(doc, "לאן נוטה: **נתמך בראיות** — תמליל (נספח 19, \"ועדה פוליטית\") + הצוות המקצועי המליץ לאשר (נספח 17). מנגד — בסיס מהותי לא-NIMBY (נספחים 25, 30).")
+
+    H2(doc, "ממצאי ניתוח-העומק — מה התחדש לאחר מיצוי 22 נספחי הרקע")
+    BULLET(doc, "**6.9 (ליבת התיק) נחלש לוועדה** — ועדת המשנה לעררים כבר אישרה הבשלה פתוחה.")
+    BULLET(doc, "**טעות הסופר בתכסית התחזקה** — שתי ערכאות הורו לתקן את חישוב השטח.")
+    BULLET(doc, "**פקיעה — הוכרע עובדתית** — סעיף הפקיעה היה בנוסח המופקד (נספח 12) והושמט מהמאושר; נותרה מחלוקת משפטית בלבד (נקר/לויתן מול חמדת הגליל).")
+    BULLET(doc, "**אופי המתקן מטה למשיבים** — תת\"ל 220 + ויתור על 8 מבני קומפוסט לטובת מתקן תרמי (נספחים 25, 30) → טיעון \"פריסת סלאמי\"/עקיפת תכנון.")
+    BULLET(doc, "**תיקון עובדתי** — פסה\"ד שדחה את העתירה (נספח 10) ניתן נגד תכנית דרך הגישה, לא נגד מי/1030.")
+    BULLET(doc, "**פער תחבורתי** — 80-92 משאיות/יום (מוסדות התכנון) מול 320-400 (ד\"ר לינק) — טעון יישוב.")
+
+    H2(doc, "פסיקה הדורשת אימות חיצוני (אינה בקורפוס הסמכותי)")
+    BULLET(doc, "ע\"א 3213/97 **נקר** — \"הדין המתפרסם ברבים מחייב\" (עוגן הוועדה לדחיית הפקיעה).")
+    BULLET(doc, "עע\"מ 4768/22 **חמדת הגליל** — פקיעת תכנית (עוגן המשיבים).")
+    BULLET(doc, "ע\"א 482/99 בלפוריה; בג\"ץ 5636/13 מתיישבי תימורים; בג\"ץ 9098/01 גניס; דנ\"א 3993/07 איקאפוד.")
+
+    H2(doc, "שאלות פתוחות להכרעת היו\"ר + סדר דיון מומלץ")
+    BULLET(doc, "(1) סיווג ההבשלה — טיפול (6.9) או אחסון (4.1.1)?")
+    BULLET(doc, "(2) דין הפקיעה לאור ההשמטה המוכחת מהנוסח המאושר.")
+    BULLET(doc, "(3) האם התכסית 3.5% היא טעות סופר הניתנת לתיקון פרשני?")
+    BULLET(doc, "(4) האם תת\"ל 220 + הוויתור על מבני הקומפוסט הופכים את הבקשה ל\"עקיפת תכנון\" (אירוס הגלבוע)?")
+    BULLET(doc, "(5) הסעד: סירוב מלא / אישור מותנה בהתאמה לשלב א' / החזרה לוועדה המקומית עם הנחיות.")
+    P(doc, "**סדר דיון מומלץ:** טענות סף (פקיעה → מיצוי/השתק → עמידה/פרסום) ← סוגיה 1 (שלביות) ← סוגיה 2 (6.9) ← סוגיה 3 (הקלות/טעות סופר) ← סוגיה 4 (תצורת בינוי) ← סוגיה 5 (שיקולים זרים).")
+
+    H2(doc, "הערכת תרחישים")
+    P(doc, "התמונה שקולה. לטובת הוועדה: סוגיה 1 (שלביות) וטיעון \"עקיפת תכנון/סלאמי\" נוכח תת\"ל 220 — חזקים. לטובת העוררים: סוגיה 2 (6.9) נחלשה, טעות הסופר התחזקה, ו-NIMBY נתמך בראיות. **התרחיש הסביר ביותר:** קבלה חלקית / החזרה מותנית או דחייה — תלוי בעיקר בשאלת היקף הבינוי בשלב א' ובשאלת \"עקיפת התכנון\". ההכרעה במובהק של יו\"ר הוועדה.")
+
+    OUT.parent.mkdir(parents=True, exist_ok=True)
+    doc.save(str(OUT))
+    print(f"saved: {OUT}")
+
+
+if __name__ == "__main__":
+    main()

scripts/final_learning_pipeline.py

@@ -91,15 +91,22 @@ async def main(args: argparse.Namespace) -> int:
 
     # The 3 steps as durable nodes (X16 / INV-DUR1) — shared runtime with
     # final_halacha (scripts/_pipeline_runtime.py). A crash/OOM in the long style
-    # panel [3] resumes from [3] instead of re-paying the Opus distillation [1].
+    # panel [3] resumes from [3] instead of re-paying the Opus distillation [2].
+    #
+    # Order (#159): enroll FIRST. enroll creates the style_corpus row (fast, no LLM)
+    # that decision_lessons attach to. Running it before the up-to-30-min distillation
+    # means the corpus exists within seconds — so the curator agent's §A (which fires
+    # in parallel on a final_learning_* wake) can record source='curator' findings
+    # without racing the 30-min step. enroll/ingest are mutually independent; panel
+    # needs both, so it stays last.
 
     async def step_ingest(results: dict) -> dict:
-        # [1] distillation (Opus) — skip if already analyzed (idempotent; --force to redo).
+        # [2] distillation (Opus) — skip if already analyzed (idempotent; --force to redo).
         status = await _latest_pair_status(case["id"])
         if status == "analyzed" and not args.force:
-            print("[1/3] ingest_final_version — דולג (הזוג כבר analyzed; --force לחידוש)")
+            print("[2/3] ingest_final_version — דולג (הזוג כבר analyzed; --force לחידוש)")
             return {"ingest": "skipped:analyzed"}
-        print("[1/3] ingest_final_version — דיסטילציית טיוטה↔סופי…", flush=True)
+        print("[2/3] ingest_final_version — דיסטילציית טיוטה↔סופי…", flush=True)
         raw = await ingest_final_version(case_number, file_path=final_path)
         try:
             env = json.loads(raw)
@@ -120,8 +127,9 @@ async def main(args: argparse.Namespace) -> int:
         return {"ingest": "done"}
 
     async def step_enroll(results: dict) -> dict:
-        # [2] enroll into style_corpus (idempotent) — lessons need a corpus_id.
+        # [1] enroll into style_corpus (idempotent) — lessons need a corpus_id, and
-        print("[2/3] רישום לקורפוס-הסגנון (idempotent)…", flush=True)
+        # it must exist early (before the long distillation) so curator §A can attach.
+        print("[1/3] רישום לקורפוס-הסגנון (idempotent)…", flush=True)
         if await _has_style_corpus(case_number):
             print("  ✓ כבר רשום בקורפוס-הסגנון")
             return {"enroll": "exists"}
@@ -150,8 +158,8 @@ async def main(args: argparse.Namespace) -> int:
         return {"panel_rc": rc or 0}
 
     steps = [
-        _pipeline_runtime.Step("ingest_final_version", step_ingest),
         _pipeline_runtime.Step("enroll_style_corpus", step_enroll),
+        _pipeline_runtime.Step("ingest_final_version", step_ingest),
         _pipeline_runtime.Step("style_panel", step_panel),
     ]
     checkpoint_db = config.DATA_DIR / "checkpoints" / "learning.sqlite"

scripts/ingest_incoming_batch.py

@@ -0,0 +1,102 @@
+"""Batch ingest of appeals-committee decisions staged in data/precedents/incoming/.
+
+Sequential (NOT concurrent — avoids the 2026-05-31 load-spike incident) ingest of
+each .doc/.docx via the canonical internal pipeline, followed by metadata extraction
+per case (the internal path does NOT auto-queue metadata — INV-ING3). Halacha is
+auto-queued by ingest; drain it separately via MCP precedent_process_pending.
+
+case_number canonical follows the filename/Nevo convention validated against the
+corpus + missing_precedents list:
+  - מרכז/חיפה/ת"א committees number with month:  NNNN/MM/YY  → NNNN-MM-YY
+  - ירושלים/צפון committees number without month: NNNN/YY     → NNNN-YY
+decision_date / summary / subject_tags / appeal_subtype are left empty on purpose —
+the metadata extractor fills them from the full text (more reliable than parsing here).
+
+Run:  mcp-server/.venv/bin/python scripts/ingest_incoming_batch.py
+Config (POSTGRES_URL, VOYAGE_API_KEY, ANTHROPIC_API_KEY) auto-loads from ~/.env.
+"""
+
+import asyncio
+import os
+import sys
+import traceback
+from pathlib import Path
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "mcp-server", "src"))
+
+from legal_mcp.services import internal_decisions as svc
+from legal_mcp.services import precedent_metadata_extractor as meta
+
+INC = "/home/chaim/legal-ai/data/precedents/incoming"
+
+# file, case_number(canonical-with-slashes), chair, district, court, practice_area
+DECISIONS = [
+    ("105-07.doc",            "105/07",       "דרור לביא-אפרת",          "צפון",     "rishuy_uvniya"),
+    ("ARAR-17-105-44.doc",    "105/17",       "רונית אלפר",              "מרכז",     "betterment_levy"),
+    ("ARAR-18-1029.doc",      "1029/18",      "אליעד וינשל",             "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-20-1018-44.doc",   "1018/20",      "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-20-1023-55.doc",   "1023/20",      "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-21-1080-55.doc",   "1080/21",      "נילי בן משה ידגר",        "צפון",     "rishuy_uvniya"),
+    ("ARAR-21-11-1051.doc",   "1051/11/21",   "רונית אלפר",              "מרכז",     "rishuy_uvniya"),
+    ("ARAR-22-01-1015.doc",   "1015/01/22",   "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-22-06-1029.doc",   "1029/06/22",   "מאיה אשכנזי",             "מרכז",     "rishuy_uvniya"),
+    ("ARAR-22-08-1044.doc",   "1044/08/22",   "מאיה אשכנזי",             "מרכז",     "rishuy_uvniya"),
+    ("ARAR-22-10-1050.doc",   "1050/10/22",   "מאיה אשכנזי",             "מרכז",     "rishuy_uvniya"),
+    ("ARAR-22-1079.doc",      "1079/22",      "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-23-04-1010.doc",   "1010/04/23",   "מאיה אשכנזי",             "מרכז",     "rishuy_uvniya"),
+    ("ARAR-23-08-1074-9.doc", "1074/08/23",   "מיכל הלברשטם דגני",       "חיפה",     "rishuy_uvniya"),
+    ("ARAR-23-1034.docx",     "1006/23",      "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-23-1073.doc",      "1073/23",      "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-23-1085.doc",      "1085/23",      "נילי בן משה ידגר",        "צפון",     "rishuy_uvniya"),
+    ("ARAR-24-01-1009-5.docx","1009/01/24",   "מיכל דגני הלברשטם",       "תל אביב",  "rishuy_uvniya"),
+    ("ARAR-24-05-1044.doc",   "1044/05/24",   "שרית אריאלי בן שמחון",    "ירושלים",  "rishuy_uvniya"),
+    ("ARAR-25-01-10072.docx", "1007/01/25",   "יפעת בן אריה שטיינברג",   "תל אביב",  "rishuy_uvniya"),
+]
+
+
+async def main():
+    results = []
+    for fname, case_number, chair, district, parea in DECISIONS:
+        fp = Path(INC) / fname
+        rec = {"file": fname, "case_number": case_number}
+        if not fp.exists():
+            rec["error"] = "file-missing"
+            print(f"✗ {fname}: file missing", flush=True)
+            results.append(rec)
+            continue
+        try:
+            out = await svc.ingest_internal_decision(
+                file_path=fp,
+                case_number=case_number,
+                chair_name=chair,
+                district=district,
+                court=f"ועדת הערר לתכנון ובנייה — מחוז {district}",
+                practice_area=parea,
+                proceeding_type="ערר",
+                is_binding=False,
+            )
+            cid = out.get("case_law_id")
+            rec["case_law_id"] = cid
+            rec["chunks"] = out.get("chunks")
+            print(f"✓ ingest {case_number}: id={cid} chunks={out.get('chunks')}", flush=True)
+            # metadata (internal path does not auto-queue it)
+            m = await meta.extract_and_apply(cid)
+            rec["meta_status"] = m.get("status")
+            sug = m.get("suggested") or {}
+            rec["suggested_case_number"] = sug.get("case_number_clean")
+            rec["citation_formatted"] = sug.get("citation_formatted")
+            rec["meta_date"] = sug.get("decision_date_iso")
+            print(f"  meta {case_number}: {m.get('status')} | clean={sug.get('case_number_clean')} | {sug.get('citation_formatted')}", flush=True)
+        except Exception as e:
+            rec["error"] = f"{type(e).__name__}: {e}"
+            print(f"✗ {fname} ({case_number}): {e}", flush=True)
+            traceback.print_exc()
+        results.append(rec)
+
+    print("\n===SUMMARY===", flush=True)
+    for r in results:
+        print(r, flush=True)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

scripts/ocr_benchmark_mistral.py

@@ -0,0 +1,392 @@
+#!/usr/bin/env python3
+"""OCR Benchmark: Current system (PyMuPDF + Google Vision) vs Mistral OCR 4.0
+
+Usage:
+    python scripts/ocr_benchmark_mistral.py [--docs N] [--output PATH]
+
+Downloads PDFs from MinIO, calls Mistral OCR API, compares against
+already-extracted text stored in the DB, and writes a Markdown report.
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import base64
+import json
+import re
+import subprocess
+import sys
+import tempfile
+import time
+from pathlib import Path
+from typing import TypedDict
+
+import httpx
+
+# ── Config ───────────────────────────────────────────────────────────────────
+
+MISTRAL_API_KEY = "UsZjLCX30ev6pox0KgvXyuFP3ktsPYpN"
+MISTRAL_OCR_MODEL = "mistral-ocr-latest"
+MISTRAL_OCR_URL = "https://api.mistral.ai/v1/ocr"
+MINIO_ALIAS = "legalminio"
+MINIO_BUCKET = "legal-documents"
+
+# Documents to benchmark — selected for diversity:
+#  - main appeal  (40 pp, 4.4 MB, large digital doc)
+#  - permit       (9 pp,  1.7 MB, likely partially scanned)
+#  - protocol     (5 pp,  178 KB, administrative typed)
+#  - response     (12 pp, ~390 KB, digital legal brief)
+DOCS_TO_BENCHMARK = [
+    {
+        "title": "כתב ערר",
+        "minio_key": "cases/1027-04-26/documents/originals/כתב ערר - מושב נחם נ׳ ועדה מקומית בית שמש.pdf",
+        "doc_type": "appeal",
+        "pages": 40,
+    },
+    {
+        "title": "נספח 1 — היתר הבנייה",
+        "minio_key": "cases/1027-04-26/documents/originals/נספח 1 - היתר הבנייה (פורסם 17.03.26).pdf",
+        "doc_type": "permit",
+        "pages": 9,
+    },
+    {
+        "title": "נספח 14 — פרוטוקול ועדה מחוזית",
+        "minio_key": "cases/1027-04-26/documents/originals/נספח 14 - פרוטוקול דיון ועדה מחוזית - 06.06.23.pdf",
+        "doc_type": "protocol",
+        "pages": 5,
+    },
+    {
+        "title": "תגובת המשיבה 3",
+        "minio_key": "cases/1027-04-26/documents/originals/תגובת המשיבה 3 לערר ולבקשה להתליית היתר בנייה.pdf",
+        "doc_type": "response",
+        "pages": 12,
+    },
+]
+
+# ── DB access (read-only: fetch already-extracted text) ──────────────────────
+
+def _fetch_extracted_texts(case_number: str) -> dict[str, str]:
+    """Pull extracted_text from DB for all docs in the case."""
+    import psycopg2  # type: ignore
+    conn = psycopg2.connect(
+        host="localhost", port=5433, dbname="legal_ai",
+        user="legal_ai", password="od0ASJZFYibOlWK59krLvvETmgqwlXe8",
+    )
+    cur = conn.cursor()
+    cur.execute(
+        """
+        SELECT d.title, d.extracted_text, d.file_path, d.page_count
+        FROM documents d
+        JOIN cases c ON c.id = d.case_id
+        WHERE c.case_number = %s AND d.extracted_text IS NOT NULL
+        """,
+        (case_number,),
+    )
+    rows = cur.fetchall()
+    cur.close()
+    conn.close()
+    return {row[0]: {"text": row[1], "file_path": row[2], "pages": row[3]} for row in rows}
+
+
+# ── MinIO download ────────────────────────────────────────────────────────────
+
+def download_from_minio(minio_key: str, dest: Path) -> None:
+    """Download a file from MinIO using mcli."""
+    src = f"{MINIO_ALIAS}/{MINIO_BUCKET}/{minio_key}"
+    result = subprocess.run(
+        ["mcli", "cp", src, str(dest)],
+        capture_output=True, text=True,
+    )
+    if result.returncode != 0:
+        raise RuntimeError(f"mcli cp failed: {result.stderr}")
+
+
+# ── Text quality metrics ─────────────────────────────────────────────────────
+
+_HEBREW_RE = re.compile(r'[֐-׿]')
+_WORD_RE = re.compile(r'\S+')
+
+
+def compute_metrics(text: str) -> dict:
+    """Compute quality metrics for extracted text."""
+    if not text:
+        return {"chars": 0, "words": 0, "avg_word_len": 0,
+                "single_char_pct": 0, "words_per_line": 0,
+                "hebrew_pct": 0, "quality_ok": False}
+
+    words = _WORD_RE.findall(text)
+    n_words = len(words)
+    if n_words == 0:
+        return {"chars": len(text), "words": 0, "avg_word_len": 0,
+                "single_char_pct": 0, "words_per_line": 0,
+                "hebrew_pct": 0, "quality_ok": False}
+
+    avg_len = sum(len(w) for w in words) / n_words
+    single_char_pct = sum(1 for w in words if len(w) == 1) / n_words
+
+    lines = [ln for ln in text.split("\n") if ln.strip()]
+    words_per_line = n_words / len(lines) if lines else 0
+
+    letters = re.findall(r'[a-zA-Z֐-׿]', text)
+    hebrew_pct = (
+        sum(1 for c in letters if _HEBREW_RE.match(c)) / len(letters)
+        if letters else 0
+    )
+
+    quality_ok = (
+        n_words >= 10
+        and avg_len >= 2.5
+        and single_char_pct <= 0.4
+        and words_per_line >= 3.0
+        and hebrew_pct >= 0.5
+    )
+
+    return {
+        "chars": len(text),
+        "words": n_words,
+        "avg_word_len": round(avg_len, 2),
+        "single_char_pct": round(single_char_pct * 100, 1),
+        "words_per_line": round(words_per_line, 1),
+        "hebrew_pct": round(hebrew_pct * 100, 1),
+        "quality_ok": quality_ok,
+    }
+
+
+def count_known_abbrev_errors(text: str) -> int:
+    """Count common Hebrew abbreviation OCR errors (pre-fix indicators)."""
+    patterns = ['עוהייד', 'עוייד', 'הנייל', 'ביהמייש', 'עייי', 'בייכ', 'תמייא']
+    return sum(text.count(p) for p in patterns)
+
+
+def count_correct_abbrevs(text: str) -> int:
+    """Count correctly rendered Hebrew abbreviations."""
+    correct = ['עו"ד', 'הנ"ל', 'ביהמ"ש', 'ע"י', 'ב"כ', 'תמ"א', 'ס"ק']
+    return sum(text.count(p) for p in correct)
+
+
+# ── Mistral OCR ───────────────────────────────────────────────────────────────
+
+def call_mistral_ocr(pdf_path: Path) -> tuple[str, float]:
+    """Call Mistral OCR API on a PDF. Returns (extracted_text, elapsed_seconds)."""
+    with open(pdf_path, "rb") as f:
+        pdf_b64 = base64.b64encode(f.read()).decode()
+
+    payload = {
+        "model": MISTRAL_OCR_MODEL,
+        "document": {
+            "type": "document_url",
+            "document_url": f"data:application/pdf;base64,{pdf_b64}",
+        },
+        "include_image_base64": False,
+    }
+
+    t0 = time.time()
+    with httpx.Client(timeout=300.0) as client:
+        resp = client.post(
+            MISTRAL_OCR_URL,
+            headers={
+                "Authorization": f"Bearer {MISTRAL_API_KEY}",
+                "Content-Type": "application/json",
+            },
+            json=payload,
+        )
+    elapsed = time.time() - t0
+
+    if resp.status_code != 200:
+        raise RuntimeError(f"Mistral OCR error {resp.status_code}: {resp.text[:500]}")
+
+    data = resp.json()
+    # Response: {"pages": [{"index": 0, "markdown": "..."}, ...]}
+    pages = data.get("pages", [])
+    text = "\n\n".join(p.get("markdown", "") for p in pages)
+    return text, elapsed
+
+
+# ── Report ────────────────────────────────────────────────────────────────────
+
+def render_metric_table(current: dict, mistral: dict) -> str:
+    rows = [
+        ("תווים", f"{current['chars']:,}", f"{mistral['chars']:,}"),
+        ("מילים", f"{current['words']:,}", f"{mistral['words']:,}"),
+        ("אורך מילה ממוצע", str(current['avg_word_len']), str(mistral['avg_word_len'])),
+        ("% מילים חד-תוויות", f"{current['single_char_pct']}%", f"{mistral['single_char_pct']}%"),
+        ("מילים לשורה", str(current['words_per_line']), str(mistral['words_per_line'])),
+        ("% תווים עבריים", f"{current['hebrew_pct']}%", f"{mistral['hebrew_pct']}%"),
+        ("איכות כוללת", "✅" if current['quality_ok'] else "❌", "✅" if mistral['quality_ok'] else "❌"),
+    ]
+    lines = ["| מדד | OCR נוכחי | Mistral OCR |", "|-----|-----------|-------------|"]
+    for label, cur_val, mis_val in rows:
+        lines.append(f"| {label} | {cur_val} | {mis_val} |")
+    return "\n".join(lines)
+
+
+def build_report(results: list[dict], output_path: Path) -> None:
+    lines = [
+        "# השוואת OCR: מערכת נוכחית מול Mistral OCR",
+        f"\n**תיק:** 1027-04-26 — בל\"מ מושב נחם מפעל בטון בית שמש  ",
+        f"**תאריך:** {time.strftime('%Y-%m-%d %H:%M')}  ",
+        f"**מודל Mistral:** `{MISTRAL_OCR_MODEL}`  ",
+        f"**מערכת נוכחית:** PyMuPDF (born-digital) + Google Cloud Vision (scanned)  ",
+        "\n---\n",
+        "## סיכום מנהלים\n",
+    ]
+
+    # Summary table
+    sum_lines = ["| מסמך | עמודים | נוכחי תווים | Mistral תווים | Mistral זמן | עדיפות |",
+                 "|------|--------|-------------|---------------|-------------|--------|"]
+    for r in results:
+        if "error" in r:
+            sum_lines.append(f"| {r['title']} | {r['pages']} | — | שגיאה | — | — |")
+            continue
+        cur_chars = r["current_metrics"]["chars"]
+        mis_chars = r["mistral_metrics"]["chars"]
+        winner = "🔵 נוכחי" if cur_chars > mis_chars * 1.05 else (
+            "🟢 Mistral" if mis_chars > cur_chars * 1.05 else "⚖️ שקול")
+        sum_lines.append(
+            f"| {r['title']} | {r['pages']} | {cur_chars:,} | {mis_chars:,} | "
+            f"{r['mistral_elapsed']:.1f}s | {winner} |"
+        )
+    lines.extend(sum_lines)
+    lines.append("\n---\n")
+
+    # Per-document detail
+    for r in results:
+        lines.append(f"## {r['title']} ({r['pages']} עמודים)\n")
+        if "error" in r:
+            lines.append(f"**שגיאה ב-Mistral OCR:** `{r['error']}`\n")
+            continue
+
+        lines.append(render_metric_table(r["current_metrics"], r["mistral_metrics"]))
+        lines.append("")
+
+        cur_abbr_err = count_known_abbrev_errors(r["current_text"])
+        mis_abbr_err = count_known_abbrev_errors(r["mistral_text"])
+        cur_abbr_ok = count_correct_abbrevs(r["current_text"])
+        mis_abbr_ok = count_correct_abbrevs(r["mistral_text"])
+
+        lines.append(f"\n**קיצורים עבריים:**")
+        lines.append(f"- נוכחי: {cur_abbr_ok} נכונים, {cur_abbr_err} שגויים")
+        lines.append(f"- Mistral: {mis_abbr_ok} נכונים, {mis_abbr_err} שגויים")
+        lines.append(f"\n**זמן Mistral:** {r['mistral_elapsed']:.1f} שניות\n")
+
+        # Side-by-side first 600 chars
+        lines.append("### דוגמת טקסט — 600 תווים ראשונים\n")
+        lines.append("**מערכת נוכחית:**")
+        lines.append("```")
+        lines.append((r["current_text"] or "")[:600].replace("```", "'''"))
+        lines.append("```\n")
+        lines.append("**Mistral OCR:**")
+        lines.append("```")
+        lines.append((r["mistral_text"] or "")[:600].replace("```", "'''"))
+        lines.append("```\n")
+        lines.append("---\n")
+
+    output_path.write_text("\n".join(lines), encoding="utf-8")
+    print(f"\n✅ דוח נשמר: {output_path}")
+
+
+# ── Main ──────────────────────────────────────────────────────────────────────
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="OCR benchmark: current vs Mistral")
+    parser.add_argument("--docs", type=int, default=4,
+                        help="כמה מסמכים לבדוק (ברירת מחדל: 4)")
+    parser.add_argument("--output", type=str,
+                        default="/home/chaim/legal-ai/data/audit/ocr-benchmark-mistral.md",
+                        help="נתיב לדוח הפלט")
+    args = parser.parse_args()
+
+    docs = DOCS_TO_BENCHMARK[: args.docs]
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    print(f"🔍 שולף טקסטים קיימים מה-DB עבור תיק 1027-04-26...")
+    db_texts = _fetch_extracted_texts("1027-04-26")
+    print(f"   נמצאו {len(db_texts)} מסמכים עם טקסט מחולץ")
+
+    results = []
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        for doc in docs:
+            print(f"\n📄 מעבד: {doc['title']} ({doc['pages']} עמודים)")
+
+            # --- Current OCR text from DB ---
+            current_text = ""
+            for title, info in db_texts.items():
+                if doc["title"].split("—")[0].strip() in title or doc["minio_key"].split("/")[-1] in (info.get("file_path") or ""):
+                    current_text = info["text"] or ""
+                    break
+            if not current_text:
+                # Try by file_path suffix match
+                key_name = doc["minio_key"].split("/")[-1]
+                for title, info in db_texts.items():
+                    if key_name in (info.get("file_path") or ""):
+                        current_text = info["text"] or ""
+                        break
+
+            if not current_text:
+                print(f"   ⚠️  לא נמצא טקסט ב-DB, מחפש לפי שם מסמך...")
+                # fallback: match by doc_type + rough title
+                for title, info in db_texts.items():
+                    if doc["doc_type"] in title.lower() or doc["title"][:8] in title:
+                        current_text = info["text"] or ""
+                        break
+
+            print(f"   נוכחי: {len(current_text):,} תווים")
+
+            # --- Download PDF from MinIO ---
+            pdf_name = doc["minio_key"].split("/")[-1]
+            pdf_path = Path(tmp_dir) / pdf_name
+            print(f"   מוריד מ-MinIO...")
+            try:
+                download_from_minio(doc["minio_key"], pdf_path)
+                print(f"   הורד: {pdf_path.stat().st_size:,} bytes")
+            except Exception as e:
+                print(f"   ❌ שגיאה בהורדה: {e}")
+                results.append({"title": doc["title"], "pages": doc["pages"], "error": str(e)})
+                continue
+
+            # --- Mistral OCR ---
+            print(f"   קורא Mistral OCR API...")
+            try:
+                mistral_text, elapsed = call_mistral_ocr(pdf_path)
+                print(f"   Mistral: {len(mistral_text):,} תווים ({elapsed:.1f}s)")
+            except Exception as e:
+                print(f"   ❌ שגיאת Mistral API: {e}")
+                results.append({
+                    "title": doc["title"], "pages": doc["pages"],
+                    "current_text": current_text,
+                    "current_metrics": compute_metrics(current_text),
+                    "mistral_text": "", "mistral_metrics": compute_metrics(""),
+                    "mistral_elapsed": 0, "error": str(e),
+                })
+                continue
+
+            results.append({
+                "title": doc["title"],
+                "pages": doc["pages"],
+                "current_text": current_text,
+                "current_metrics": compute_metrics(current_text),
+                "mistral_text": mistral_text,
+                "mistral_metrics": compute_metrics(mistral_text),
+                "mistral_elapsed": elapsed,
+            })
+
+    print("\n📊 בונה דוח...")
+    build_report(results, output_path)
+
+    # Also save raw texts for manual inspection
+    raw_dir = output_path.parent / "ocr-benchmark-raw"
+    raw_dir.mkdir(exist_ok=True)
+    for r in results:
+        safe = r["title"].replace("/", "-").replace(" ", "_")[:40]
+        if "current_text" in r:
+            (raw_dir / f"{safe}_current.txt").write_text(r["current_text"], encoding="utf-8")
+        if "mistral_text" in r:
+            (raw_dir / f"{safe}_mistral.txt").write_text(r["mistral_text"], encoding="utf-8")
+    print(f"💾 טקסטים גולמיים נשמרו: {raw_dir}")
+
+
+if __name__ == "__main__":
+    main()

scripts/renumber_cases.py

@@ -19,7 +19,13 @@ migrates atomically per case — is everything that embeds the number as *text*:
   4. MinIO keys cases/{old}/…       3 buckets; cp→new then rm old.
      legal-immutable (WORM/object-lock) → copy-only, old object stays locked.
   5. Gitea repo cases/{old}         API PATCH name + local .git remote rewrite
-  6. Paperclip project name         replace(old→new) so case↔project lookup holds
+  6. Paperclip case↔issue linkage   replace(old→new) in THREE places, because the
+     legal-ai → Paperclip lookup (get_case_issues) keys on the case number as text:
+       • projects.name                 so case↔project lookup holds
+       • plugin_state.legal-case-number the authoritative issue linkage value_json
+       • issues.title                   the '[ערר {cn}] …' tag the title-path lookup uses
+     Without (b)+(c) the issues keep the OLD number and get_case_issues returns [],
+     so post-final actions (run-learning / run-halacha) silently skip ("no_issue").
 
 Bare occurrences of the old number that are NOT inside a '/cases/{old}/' path
 (e.g. prose in notes, a citation) are *reported for review*, never auto-edited.
@@ -273,10 +279,25 @@ async def inspect_paperclip(old: str) -> dict:
     try:
         c = await asyncpg.connect(PAPERCLIP_DSN, timeout=10)
     except Exception as e:
-        return {"reachable": False, "error": str(e)[:120], "projects": []}
+        return {"reachable": False, "error": str(e)[:120],
+                "projects": [], "linkage_rows": 0, "issue_titles": 0}
     try:
         rows = await c.fetch("SELECT id, name FROM projects WHERE name LIKE $1", f"%{old}%")
-        return {"reachable": True, "projects": [(str(r["id"]), r["name"]) for r in rows]}
+        # The two surfaces get_case_issues actually keys on: the legal-case-number
+        # plugin_state linkage and the '[ערר {cn}]' tag in issue titles.
+        linkage = await c.fetchval(
+            "SELECT count(*) FROM plugin_state "
+            "WHERE state_key = 'legal-case-number' AND value_json = to_jsonb($1::text)",
+            old,
+        )
+        titles = await c.fetchval(
+            "SELECT count(*) FROM issues WHERE title LIKE $1", f"%{old}%")
+        return {
+            "reachable": True,
+            "projects": [(str(r["id"]), r["name"]) for r in rows],
+            "linkage_rows": linkage or 0,
+            "issue_titles": titles or 0,
+        }
     finally:
         await c.close()
 
@@ -369,16 +390,31 @@ async def apply_case(conn, rec: dict, *, skip_minio: bool, skip_gitea: bool,
         except urllib.error.HTTPError as e:
             log(f"  ✗ Gitea rename failed: HTTP {e.code} {e.read()[:160]!r}")
 
-    # 6. Paperclip project name
+    # 6. Paperclip case↔issue linkage — project name + legal-case-number value +
-    if not skip_paperclip and rec["paperclip"].get("reachable") and rec["paperclip"]["projects"]:
+    # issue titles. (b)+(c) are what get_case_issues keys on; without them the
+    # issues keep the old number and run-learning/run-halacha skip with "no_issue".
+    if not skip_paperclip and rec["paperclip"].get("reachable"):
         import asyncpg
         c = await asyncpg.connect(PAPERCLIP_DSN, timeout=10)
         try:
+            if rec["paperclip"]["projects"]:
+                res = await c.execute(
+                    "UPDATE projects SET name = replace(name, $1, $2), updated_at = now() "
+                    "WHERE name LIKE $3",
+                    old, new, f"%{old}%",
+                )
+                log(f"  ✓ Paperclip projects: {res}")
             res = await c.execute(
-                "UPDATE projects SET name = replace(name, $1, $2), updated_at = now() WHERE name LIKE $3",
+                "UPDATE plugin_state SET value_json = to_jsonb($2::text) "
+                "WHERE state_key = 'legal-case-number' AND value_json = to_jsonb($1::text)",
+                old, new,
+            )
+            log(f"  ✓ Paperclip plugin_state (legal-case-number): {res}")
+            res = await c.execute(
+                "UPDATE issues SET title = replace(title, $1, $2) WHERE title LIKE $3",
                 old, new, f"%{old}%",
             )
-            log(f"  ✓ Paperclip projects: {res}")
+            log(f"  ✓ Paperclip issue titles: {res}")
         finally:
             await c.close()
 
@@ -419,6 +455,8 @@ def print_inspection(rec: dict) -> None:
             log(f"     pclip:  {name}")
         if not pc["projects"]:
             log("     pclip:  (no matching project)")
+        log(f"     pclip:  linkage rows={pc.get('linkage_rows', 0)}  "
+            f"issue titles={pc.get('issue_titles', 0)}  (→ rewritten to {rec['new']})")
     else:
         log(f"     pclip:  unreachable ({pc.get('error','')})")
     log("     DB path columns to rewrite:")

scripts/style_lesson_panel.py

@@ -110,6 +110,19 @@ def _category(change: dict) -> str:
     return "style"
 
 
+# Graduated gate (INV-LRN1, chair decision 2026-06-28): a STYLE lesson the panel
+# kept by 2/2 consensus flows straight to the writer (review_status='approved'),
+# reversibly — the chair can veto it in /training. SUBSTANCE (halacha/precedent/
+# fact) never reaches here (it's filtered to `substance` and skipped, and routes
+# through the strict 3-judge halacha gate), so every category this panel emits is
+# style and auto-approves. The constant keeps the gate explicit and future-proof.
+_STYLE_CATEGORIES = frozenset({"style", "structure", "lexicon", "tabular"})
+
+
+def _review_status_for(category: str) -> str:
+    return "approved" if category in _STYLE_CATEGORIES else "proposed"
+
+
 # ── two judges, one signature: (system, user) -> dict|None ──
 
 async def judge_deepseek(client: httpx.AsyncClient, system: str, user: str) -> dict | None:
@@ -323,18 +336,24 @@ async def main(args: argparse.Namespace) -> int:
                         _lesson_text(r["_change"])])
 
     written = 0
+    approved = 0
     for r in fresh:
+        cat = _category(r["_change"])
+        rs = _review_status_for(cat)
         await db.add_decision_lesson(
             UUID(corpus_id),
             lesson_text=_lesson_text(r["_change"]),
-            category=_category(r["_change"]),
+            category=cat,
             source="panel:deepseek+gemini",
             created_by="panel",
+            review_status=rs,
         )
         written += 1
+        approved += (rs == "approved")
 
     chair = cc["split"] + cc["incomplete"]
-    print(f"\nAPPLIED (reversible): wrote {written} decision_lesson proposals "
+    print(f"\nAPPLIED (reversible): wrote {written} decision_lessons "
+          f"({approved} auto-approved style → writer; graduated gate) "
           f"(source=panel:deepseek+gemini) · {skipped_dup} כפילויות דולגו · "
           f"{chair} escalated to chair · {len(substance)} substance skipped")
     print(f"backup → {backup}")

skills/docx/references/advanced-features.md

@@ -0,0 +1,170 @@
+# Advanced Features — פיצ'רים מתקדמים
+
+## הערות שוליים (Footnotes)
+
+**שימוש מרכזי:** הפניות לחקיקה ופסיקה.
+
+```javascript
+const { FootnoteReferenceRun } = require('docx');
+
+const doc = new Document({
+  footnotes: {
+    1: { children: [new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      children: [new TextRun({
+        text: "חוק החוזים (חלק כללי), התשל״ג-1973, סעיף 12.",
+        font: "David", size: 20, rightToLeft: true
+      })]
+    })] },
+  },
+  // ...sections
+});
+
+// הפניה בגוף הטקסט:
+new Paragraph({
+  bidirectional: true, alignment: AlignmentType.BOTH,
+  children: [
+    new TextRun({ text: "חובת תום הלב", font: "David", size: 24, rightToLeft: true }),
+    new FootnoteReferenceRun(1),
+    new TextRun({ text: " חלה על כל שלבי המשא ומתן.", font: "David", size: 24, rightToLeft: true }),
+  ]
+})
+```
+
+### תיקון RTL בהערות שוליים (post-unpack)
+docx-js לא מגדיר RTL מלא. אחרי unpack, תקן ב-`word/footnotes.xml`:
+```xml
+<w:footnote w:id="1">
+  <w:p>
+    <w:pPr>
+      <w:pStyle w:val="FootnoteText"/>
+      <w:bidi/>
+      <w:jc w:val="start"/>
+    </w:pPr>
+    <w:r>
+      <w:rPr>
+        <w:rStyle w:val="FootnoteReference"/>
+        <w:rtl/>
+      </w:rPr>
+      <w:footnoteRef/>
+    </w:r>
+  </w:p>
+</w:footnote>
+```
+
+---
+
+## תוכן עניינים (TOC)
+
+**⚠️ TOC ידני בלבד** — `TableOfContents` של docx-js מאבד הגדרות RTL בעדכון Word.
+
+```javascript
+const { Tab, TabStopType, LeaderType, LineRuleType } = require('docx');
+
+const tocEntry = (text, pageNum, opts = {}) => new Paragraph({
+  bidirectional: true,
+  spacing: { after: 60, line: 276, lineRule: LineRuleType.AUTO },
+  ...(opts.indent ? { indent: { right: opts.indent } } : {}),
+  tabStops: [{ type: TabStopType.RIGHT, position: 9026, leader: LeaderType.DOT }],
+  children: [
+    new TextRun({ text, font: "David", size: 24, rightToLeft: true, bold: opts.bold || false }),
+    new TextRun({ children: [new Tab()], font: "David", rightToLeft: true }),
+    new TextRun({ text: String(pageNum), font: "David", size: 24, rightToLeft: true }),
+  ]
+});
+
+// שימוש:
+tocEntry("פרק א׳ — הגדרות כלליות", 2, { bold: true }),
+tocEntry("1. הגדרות יסוד", 2, { indent: 400 }),
+```
+
+---
+
+## מספר סקשנים (Multiple Sections)
+
+**שימוש:** כותרות שונות לנספחים, שוליים שונים.
+
+```javascript
+const doc = new Document({
+  sections: [
+    {
+      properties: {
+        page: { size: { width: 11906, height: 16838 }, margin: { top: 1417, right: 1417, bottom: 1417, left: 1417 } },
+        bidi: true,
+      },
+      headers: {
+        default: new Header({ children: [new Paragraph({
+          bidirectional: true, alignment: AlignmentType.CENTER,
+          children: [new TextRun({ text: "הסכם שירותים", font: "David", size: 20, bold: true, rightToLeft: true })]
+        })] })
+      },
+      children: [ /* גוף ההסכם */ ]
+    },
+    {
+      properties: {
+        page: { size: { width: 11906, height: 16838 }, margin: { top: 1417, right: 1417, bottom: 1417, left: 1417 } },
+        bidi: true,
+      },
+      headers: {
+        default: new Header({ children: [new Paragraph({
+          bidirectional: true, alignment: AlignmentType.START,
+          children: [new TextRun({ text: "נספח א׳ — לוח תעריפים", font: "David", size: 20, bold: true, rightToLeft: true })]
+        })] })
+      },
+      children: [ /* הנספח */ ]
+    }
+  ]
+});
+```
+
+---
+
+## לוגו/תמונה בכותרת (Letterhead)
+
+```javascript
+const { ImageRun } = require('docx');
+const logoBuffer = fs.readFileSync('/path/to/logo.png');
+
+headers: {
+  default: new Header({
+    children: [
+      new Paragraph({
+        alignment: AlignmentType.CENTER,
+        children: [new ImageRun({ data: logoBuffer, transformation: { width: 200, height: 60 }, type: "png" })],
+      }),
+      new Paragraph({
+        bidirectional: true, alignment: AlignmentType.CENTER,
+        children: [new TextRun({ text: "משרד עורכי דין ישראלי ושות׳", font: "David", size: 20, bold: true, rightToLeft: true })],
+      }),
+    ],
+  }),
+}
+```
+
+**הערה:** תמונה חייבת להיות קובץ אמיתי — לבקש מהמשתמש אם אין.
+
+---
+
+## היפרלינקים
+
+```javascript
+const { ExternalHyperlink, UnderlineType } = require('docx');
+
+new Paragraph({
+  bidirectional: true,
+  children: [
+    new TextRun({ text: "ראה: ", font: "David", size: 24, rightToLeft: true }),
+    new ExternalHyperlink({
+      link: "https://www.nevo.co.il/law_html/law01/073_002.htm",
+      children: [new TextRun({
+        text: "חוק החוזים באתר נבו",
+        font: "David", size: 24, rightToLeft: true,
+        color: "0563C1",
+        underline: { type: UnderlineType.SINGLE },
+      })],
+    }),
+  ]
+})
+```
+
+**⚠️ אל תשתמש ב-`style: "Hyperlink"`** — מפריע ל-RTL. הגדר `color` + `underline` ידנית.

skills/docx/references/document-templates.md

@@ -0,0 +1,219 @@
+# Document Templates — תבניות מסמכים משפטיים
+
+## תבנית 1: כתב טענות (בקשה, תביעה, הגנה, ערעור)
+
+```javascript
+const { Document, Packer, Paragraph, TextRun, Table, TableRow, TableCell,
+        AlignmentType, LevelFormat, BorderStyle, WidthType } = require('docx');
+
+const PAGE_WIDTH = 11906;
+const MARGINS = { top: 1134, right: 1134, bottom: 1134, left: 1134 };
+const CONTENT_WIDTH = PAGE_WIDTH - MARGINS.left - MARGINS.right;
+
+const noBorder = { style: BorderStyle.NONE, size: 0, color: "FFFFFF" };
+const noBorders = { top: noBorder, bottom: noBorder, left: noBorder, right: noBorder };
+
+// Header בית משפט — טבלה עם שם בית המשפט (ימין) ומספר תיק (שמאל)
+function courtHeader(courtName, caseNumber) {
+  return new Table({
+    width: { size: CONTENT_WIDTH, type: WidthType.DXA },
+    columnWidths: [CONTENT_WIDTH / 2, CONTENT_WIDTH / 2],
+    visuallyRightToLeft: true,
+    rows: [
+      new TableRow({
+        children: [
+          new TableCell({
+            width: { size: CONTENT_WIDTH / 2, type: WidthType.DXA },
+            borders: noBorders,
+            children: [new Paragraph({
+              bidirectional: true, alignment: AlignmentType.START,
+              children: [new TextRun({ text: courtName, bold: true, font: "David", size: 26, rightToLeft: true })]
+            })]
+          }),
+          new TableCell({
+            width: { size: CONTENT_WIDTH / 2, type: WidthType.DXA },
+            borders: noBorders,
+            children: [new Paragraph({
+              bidirectional: true, alignment: AlignmentType.END,
+              children: [new TextRun({ text: caseNumber, bold: true, font: "David", size: 26, rightToLeft: true })]
+            })]
+          })
+        ]
+      })
+    ]
+  });
+}
+
+function mainTitle(text) {
+  return new Paragraph({
+    bidirectional: true, alignment: AlignmentType.CENTER,
+    spacing: { before: 300, after: 300 },
+    children: [new TextRun({ text, bold: true, font: "David", size: 28, rightToLeft: true, underline: {} })]
+  });
+}
+
+function subHeading(text) {
+  return new Paragraph({
+    bidirectional: true, alignment: AlignmentType.START,
+    spacing: { before: 240, after: 120 },
+    children: [new TextRun({ text, bold: true, font: "David", size: 24, rightToLeft: true, underline: {} })]
+  });
+}
+
+const doc = new Document({
+  numbering: {
+    config: [{
+      reference: "legal-clauses",
+      levels: [{
+        level: 0, format: LevelFormat.DECIMAL, text: "%1.",
+        alignment: AlignmentType.START, suffix: "tab",
+        style: { paragraph: { indent: { left: 360, hanging: 360 } } }
+      }]
+    }]
+  },
+  sections: [{
+    properties: {
+      page: { size: { width: PAGE_WIDTH, height: 16838 }, margin: MARGINS },
+      bidi: true
+    },
+    children: [
+      courtHeader("בית המשפט המחוזי בתל אביב", "ת\"א 12345-01-26"),
+      mainTitle("כתב תביעה"),
+      // ... פרטי צדדים, סעיפים, חתימה
+    ]
+  }]
+});
+```
+
+---
+
+## תבנית 2: מכתב התראה
+
+```javascript
+function letterHeader(firmName, address, phone, email) {
+  return [
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      children: [new TextRun({ text: firmName, bold: true, font: "David", size: 28, rightToLeft: true })]
+    }),
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      children: [new TextRun({ text: address, font: "David", size: 22, rightToLeft: true })]
+    }),
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      spacing: { after: 300 },
+      children: [new TextRun({ text: `טל': ${phone} | ${email}`, font: "David", size: 22, rightToLeft: true })]
+    }),
+  ];
+}
+
+function subjectLine(text) {
+  return new Paragraph({
+    bidirectional: true, alignment: AlignmentType.CENTER,
+    spacing: { before: 200, after: 200 },
+    children: [
+      new TextRun({ text: "הנדון: ", bold: true, font: "David", size: 24, rightToLeft: true }),
+      new TextRun({ text, bold: true, font: "David", size: 24, rightToLeft: true, underline: {} })
+    ]
+  });
+}
+
+// שימוש:
+sections: [{
+  properties: { page: { size: { width: 11906, height: 16838 }, margin: { top: 1417, right: 1417, bottom: 1417, left: 1417 } }, bidi: true },
+  children: [
+    ...letterHeader("משרד עו\"ד כהן ושות'", "רח' הרצל 1, תל אביב", "03-1234567", "office@cohen-law.co.il"),
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      children: [new TextRun({ text: "תאריך: 10.2.2026", font: "David", size: 24, rightToLeft: true })]
+    }),
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.START,
+      spacing: { before: 200 },
+      children: [new TextRun({ text: "לכבוד: [שם הנמען]", font: "David", size: 24, rightToLeft: true })]
+    }),
+    subjectLine("התראה בטרם נקיטת הליכים משפטיים"),
+  ]
+}]
+```
+
+---
+
+## תבנית 3: הסכם/חוזה
+
+```javascript
+const CONTENT_WIDTH = 9638; // A4 עם שוליים 2.5 ס"מ
+const noBorders = { /* ראה תבנית 1 */ };
+
+function contractTitle(text) {
+  return new Paragraph({
+    bidirectional: true, alignment: AlignmentType.CENTER,
+    spacing: { after: 300 },
+    children: [new TextRun({ text, bold: true, font: "David", size: 32, rightToLeft: true })]
+  });
+}
+
+function partyClause(label, name, id, address, alias) {
+  return new Paragraph({
+    bidirectional: true, alignment: AlignmentType.BOTH,
+    spacing: { after: 120 },
+    children: [
+      new TextRun({ text: `${label}: `, bold: true, font: "David", size: 24, rightToLeft: true }),
+      new TextRun({ text: `${name}, ח.פ./ת.ז. ${id}, מ${address} (להלן: "`, font: "David", size: 24, rightToLeft: true }),
+      new TextRun({ text: alias, bold: true, font: "David", size: 24, rightToLeft: true }),
+      new TextRun({ text: '")', font: "David", size: 24, rightToLeft: true }),
+    ]
+  });
+}
+
+function signatureTable(contentWidth) {
+  return new Table({
+    width: { size: contentWidth, type: WidthType.DXA },
+    columnWidths: [contentWidth / 2, contentWidth / 2],
+    visuallyRightToLeft: true,
+    rows: [new TableRow({
+      children: [
+        new TableCell({
+          borders: noBorders,
+          children: [
+            new Paragraph({ bidirectional: true, alignment: AlignmentType.CENTER,
+              children: [new TextRun({ text: "_________________", font: "David", size: 24, rightToLeft: true })] }),
+            new Paragraph({ bidirectional: true, alignment: AlignmentType.CENTER,
+              children: [new TextRun({ text: "צד א'", font: "David", size: 24, rightToLeft: true })] })
+          ]
+        }),
+        new TableCell({
+          borders: noBorders,
+          children: [
+            new Paragraph({ bidirectional: true, alignment: AlignmentType.CENTER,
+              children: [new TextRun({ text: "_________________", font: "David", size: 24, rightToLeft: true })] }),
+            new Paragraph({ bidirectional: true, alignment: AlignmentType.CENTER,
+              children: [new TextRun({ text: "צד ב'", font: "David", size: 24, rightToLeft: true })] })
+          ]
+        })
+      ]
+    })]
+  });
+}
+
+sections: [{
+  properties: { page: { size: { width: 11906, height: 16838 }, margin: { top: 1417, right: 1417, bottom: 1417, left: 1417 } }, bidi: true },
+  children: [
+    contractTitle("הסכם שירותים"),
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.CENTER,
+      children: [new TextRun({ text: "נערך ונחתם בתל אביב ביום __________", font: "David", size: 24, rightToLeft: true })]
+    }),
+    partyClause("מצד אחד", "[שם]", "[מספר]", "[כתובת]", "המזמין"),
+    partyClause("מצד שני", "[שם]", "[מספר]", "[כתובת]", "הספק"),
+    // הואילים + סעיפים...
+    new Paragraph({
+      bidirectional: true, alignment: AlignmentType.CENTER,
+      spacing: { before: 400, after: 300 },
+      children: [new TextRun({ text: "ולראיה באו הצדדים על החתום:", bold: true, font: "David", size: 24, rightToLeft: true })]
+    }),
+    signatureTable(CONTENT_WIDTH)
+  ]
+}]
+```

skills/docx/references/tracked-changes.md

@@ -0,0 +1,57 @@
+# Tracked Changes — עקוב אחר שינויים
+
+## שם מחבר בעברית
+```xml
+<w:del w:id="10" w:author="עו&quot;ד כהן" w:date="2026-02-06T09:00:00Z">
+```
+
+## שינוי ערך (סכום, תאריך, תקופה)
+פצל את הטקסט ועטוף רק את הערך שמשתנה:
+```xml
+<w:r><w:rPr>...RTL PROPS...</w:rPr>
+  <w:t xml:space="preserve">שכר הטרחה יעמוד על סך של </w:t></w:r>
+<w:del w:id="10" w:author="עו&quot;ד כהן" w:date="...">
+  <w:r><w:rPr>...RTL PROPS...</w:rPr><w:delText>750</w:delText></w:r>
+</w:del>
+<w:ins w:id="11" w:author="עו&quot;ד כהן" w:date="...">
+  <w:r><w:rPr>...RTL PROPS...</w:rPr><w:t>850</w:t></w:r>
+</w:ins>
+<w:r><w:rPr>...RTL PROPS...</w:rPr>
+  <w:t xml:space="preserve"> ש״ח לשעת עבודה</w:t></w:r>
+```
+
+## מחיקת סעיף שלם
+```xml
+<w:p>
+  <w:pPr>
+    <w:bidi/>
+    <w:jc w:val="both"/>
+    <w:rPr>
+      <w:del w:id="20" w:author="עו&quot;ד כהן" w:date="..."/>
+    </w:rPr>
+  </w:pPr>
+  <w:del w:id="21" w:author="עו&quot;ד כהן" w:date="...">
+    <w:r><w:rPr>...RTL PROPS...</w:rPr>
+      <w:delText>הסעיף שנמחק</w:delText></w:r>
+  </w:del>
+</w:p>
+```
+
+## RTL PROPS — בלוק rPr מלא לכל run
+```xml
+<w:rPr>
+  <w:rFonts w:ascii="David" w:cs="David" w:eastAsia="David" w:hAnsi="David"/>
+  <w:sz w:val="24"/>
+  <w:szCs w:val="24"/>
+  <w:rtl/>
+</w:rPr>
+```
+
+## קבלה/דחייה של שינויים
+
+| פעולה | לפני | אחרי |
+|-------|------|------|
+| קבלת הוספה | `<w:ins ...><w:r>...<w:t>טקסט</w:t></w:r></w:ins>` | `<w:r>...<w:t>טקסט</w:t></w:r>` |
+| דחיית הוספה | `<w:ins ...><w:r>...</w:r></w:ins>` | *(מחק הכל)* |
+| קבלת מחיקה | `<w:del ...><w:r>...<w:delText>טקסט</w:delText></w:r></w:del>` | *(מחק הכל)* |
+| דחיית מחיקה | `<w:del ...><w:r>...<w:delText>טקסט</w:delText></w:r></w:del>` | `<w:r>...<w:t>טקסט</w:t></w:r>` |

web-ui/src/app/cases/[caseNumber]/compose/page.tsx

@@ -9,10 +9,15 @@ import { Button } from "@/components/ui/button";
 import { Skeleton } from "@/components/ui/skeleton";
 import { SubsectionCard } from "@/components/compose/subsection-card";
 import { PrecedentsSection } from "@/components/compose/precedents-section";
+import { CitationVerificationPanel } from "@/components/compose/citation-verification-panel";
 import { DecisionBlocksPanel } from "@/components/cases/decision-blocks-panel";
 import { Markdown } from "@/components/ui/markdown";
 import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs";
 import { useCase, type CaseStatus } from "@/lib/api/cases";
+import {
+  useCaseLearningStatus,
+  type CaseLearningStatus,
+} from "@/lib/api/learning";
 import { useResearchAnalysis } from "@/lib/api/research";
 import { useCasePrecedents } from "@/lib/api/precedents";
 import { APPEAL_SUBTYPES } from "@/lib/practice-area";
@@ -51,6 +56,45 @@ function subtypeLabel(subtype?: string | null): string | null {
   return APPEAL_SUBTYPES.find((s) => s.value === subtype)?.label ?? null;
 }
 
+// ── Staged-pipeline indicator text (mockup 03) — derived from the live
+// learning-status, same source as the drafts-panel LearningStatusBadges. ──────
+function voiceLearningText(s?: CaseLearningStatus): string {
+  if (!s?.final_uploaded) return "ממתין להעלאת הסופי";
+  const v = s.voice_learning;
+  if (v.outcome === "succeeded") {
+    const bits = [`${v.lessons_count} לקחים הופקו`];
+    if (v.lessons_proposed > 0) bits.push(`${v.lessons_proposed} הוצעו לאישור`);
+    return `✓ הושלם · ${bits.join(" · ")}`;
+  }
+  if (v.outcome === "failed") return v.error ? `✗ נכשל — ${v.error}` : "✗ נכשל";
+  return "ממתין להרצה";
+}
+
+function halachaExtractionText(s?: CaseLearningStatus): string {
+  if (!s?.final_uploaded) return "ממתין להעלאת הסופי";
+  const h = s.halacha_extraction;
+  if (!h.enrolled_in_corpus)
+    return h.not_enrolled_reason ?? "לא נכנס לקורפוס-הפסיקה";
+  switch (h.status) {
+    case "completed":
+      return `✓ הושלם · חולצו ${h.halachot_count} · ${h.approved} אושרו · ${h.rejected} נדחו`;
+    case "processing":
+      return "רץ עכשיו…";
+    case "pending":
+    case "busy":
+      return "בתור";
+    case "partial":
+      return `חלקי · חולצו ${h.halachot_count}`;
+    case "failed":
+    case "extraction_failed":
+      return "✗ נכשל";
+    case "no_chunks":
+      return "אין טקסט לחילוץ";
+    default:
+      return "ממתין להרצה";
+  }
+}
+
 function ProseSection({ title, content }: { title: string; content?: string }) {
   if (!content?.trim()) return null;
   return (
@@ -76,6 +120,7 @@ function FinishRail({
   const fileRef = useRef<HTMLInputElement>(null);
   const [uploading, setUploading] = useState(false);
   const [uploadMsg, setUploadMsg] = useState<{ ok: boolean; text: string } | null>(null);
+  const learning = useCaseLearningStatus(caseNumber);
 
   async function handleUpload(file: File) {
     setUploading(true);
@@ -167,10 +212,10 @@ function FinishRail({
         {/* mockup 03: stage indicators — informational pointers, not actions */}
         <div className="mt-3 space-y-0">
           <div className="text-[0.78rem] text-ink-muted pt-2 border-t border-rule-soft">
-            <b className="text-navy">הרץ למידת-קול</b> — ממתין להעלאת הסופי
+            <b className="text-navy">הרץ למידת-קול</b> — {voiceLearningText(learning.data)}
           </div>
           <div className="text-[0.78rem] text-ink-muted pt-2 mt-2 border-t border-rule-soft">
-            <b className="text-navy">הרץ אימות-הלכות</b> — ממתין להעלאת הסופי
+            <b className="text-navy">הרץ אימות-הלכות</b> — {halachaExtractionText(learning.data)}
           </div>
         </div>
 
@@ -236,6 +281,12 @@ export default function ComposePage({
           <span aria-hidden>·</span>
           <span className="text-navy">עורך החלטה</span>
         </nav>
+        <Link
+          href={`/cases/${caseNumber}`}
+          className="inline-flex items-center gap-1.5 mb-3 rounded-md border border-rule bg-surface px-3 py-1.5 text-[0.82rem] font-semibold text-gold-deep hover:bg-gold-wash hover:text-navy transition-colors"
+        >
+          <span aria-hidden>→</span> חזרה לדף התיק
+        </Link>
         <div className="flex items-center gap-3 flex-wrap">
           <h1 className="text-navy text-2xl font-bold mb-0">ערר {caseNumber}</h1>
           <StatusChip status={caseQuery.data?.status} />
@@ -270,6 +321,7 @@ export default function ComposePage({
               <TabsList className="bg-rule-soft/60">
                 <TabsTrigger value="blocks">עורך הבלוקים</TabsTrigger>
                 <TabsTrigger value="positions">עמדות וטענות</TabsTrigger>
+                <TabsTrigger value="verify">אימות פסיקה</TabsTrigger>
               </TabsList>
 
               {/* Tab 1 — the 12-block decision editor (reused DecisionBlocksPanel) */}
@@ -277,6 +329,11 @@ export default function ComposePage({
                 <DecisionBlocksPanel caseNumber={caseNumber} />
               </TabsContent>
 
+              {/* Tab 3 — citation verification: per-argument support + verify gate (#154) */}
+              <TabsContent value="verify" className="mt-5">
+                <CitationVerificationPanel caseNumber={caseNumber} />
+              </TabsContent>
+
               {/* Tab 2 — chair positions on the analyst's threshold-claims + issues */}
               <TabsContent value="positions" className="mt-5">
                 {analysis.isPending ? (

web-ui/src/app/cases/[caseNumber]/page.tsx

@@ -39,14 +39,16 @@ export default function CaseDetailPage({
   params: Promise<{ caseNumber: string }>;
 }) {
   const { caseNumber } = use(params);
-  const { data, isPending, error } = useCase(caseNumber);
+  const { data, isPending, error, refetch } = useCase(caseNumber);
   const startWorkflow = useStartWorkflow(caseNumber);
   const canStartWorkflow = data?.status === "new" || data?.status === "documents_ready";
   const expectedOutcomeLabel = data?.expected_outcome
     ? EXPECTED_OUTCOME_LABELS[data.expected_outcome] ?? data.expected_outcome
     : null;
 
-  if (error) {
+  // Only take over the whole page when there is NO data to show. A transient
+  // 5xx on the 5s background refetch must not blow away an already-loaded page.
+  if (error && !data) {
     return (
       <AppShell>
         <section className="space-y-6">
@@ -54,9 +56,14 @@ export default function CaseDetailPage({
             <CardContent className="px-6 py-6 text-center space-y-3">
               <p className="text-danger font-semibold">שגיאה בטעינת התיק</p>
               <p className="text-sm text-ink-muted">{error.message}</p>
-              <Button asChild variant="outline">
+              <div className="flex items-center justify-center gap-2">
-                <Link href="/">חזרה לרשימת התיקים</Link>
+                <Button variant="outline" onClick={() => refetch()}>
-              </Button>
+                  נסה שוב
+                </Button>
+                <Button asChild variant="ghost">
+                  <Link href="/">חזרה לרשימת התיקים</Link>
+                </Button>
+              </div>
             </CardContent>
           </Card>
         </section>
@@ -91,6 +98,9 @@ export default function CaseDetailPage({
     <>
       {data && <CaseEditDialog data={data} />}
       <UploadSheet caseNumber={caseNumber} />
+      <Button asChild className="bg-gold text-white hover:bg-gold-deep border-transparent">
+        <Link href={`/cases/${caseNumber}/compose`}>פתח עורך החלטה</Link>
+      </Button>
       {canStartWorkflow && (
         <Button
           className="bg-gold-deep hover:bg-gold-deep/90 text-parchment"
@@ -135,16 +145,7 @@ export default function CaseDetailPage({
               <DocumentsPanel data={data} />
 
               <AgentActivityPreview caseNumber={caseNumber} />
-
+              {/* decision-editor CTA moved to the band actions (visible on all tabs) */}
-              {/* gold CTA — open the decision editor */}
-              <Button
-                asChild
-                className="w-full bg-gold text-white hover:bg-gold-deep border-transparent py-6 text-base font-semibold"
-              >
-                <Link href={`/cases/${caseNumber}/compose`}>
-                  פתח עורך החלטה →
-                </Link>
-              </Button>
             </TabsContent>
 
             <TabsContent value="arguments" className="mt-0">

web-ui/src/app/missing-precedents/page.tsx

@@ -22,12 +22,12 @@ import { MissingPrecedentsTable } from "@/components/missing-precedents/missing-
 
 type StatusFilter = MissingPrecedentStatus | "all";
 
+// Only the two states the chair acts on: open gaps to fill, and closed gaps for
+// reference. "הועלה" (transient) and "לא-רלוונטי" were dropped from the filter,
+// and "הכל" with them (chair's request, design 09).
 const STATUS_CHIPS: { value: StatusFilter; label: string }[] = [
   { value: "open", label: "פתוח" },
-  { value: "uploaded", label: "הועלה" },
   { value: "closed", label: "נסגר" },
-  { value: "irrelevant", label: "לא-רלוונטי" },
-  { value: "all", label: "הכל" },
 ];
 
 export default function MissingPrecedentsPage() {
@@ -145,15 +145,13 @@ export default function MissingPrecedentsPage() {
         <div className="rounded-lg border border-rule bg-parchment px-5 py-3.5 text-[0.82rem] text-ink-muted leading-7">
           <b className="text-ink-soft">מחזור-חיים:</b>{" "}
           <LifecycleChip tone="open">פתוח</LifecycleChip> →{" "}
-          <LifecycleChip tone="up">הועלה</LifecycleChip> →{" "}
           <LifecycleChip tone="closed">נסגר</LifecycleChip>. פריט נפתח אוטומטית
           בעת חילוץ ציטוט שאין לו תקדים בקורפוס; בהעלאת פסק-הדין הוא מקושר לרשומת
           הפסיקה דרך{" "}
           <code className="rounded border border-rule bg-surface px-1.5 py-0.5 text-[0.75rem] text-gold-deep" dir="ltr">
             linked_case_law_id
           </code>{" "}
-          ונסגר. פריט שאינו רלוונטי מסומן{" "}
+          ונסגר.
-          <LifecycleChip tone="na">לא-רלוונטי</LifecycleChip> מבלי שתידרש העלאה.
         </div>
       </section>
     </AppShell>

web-ui/src/app/operations/page.tsx

@@ -12,7 +12,6 @@ import { Switch } from "@/components/ui/switch";
 import { Input } from "@/components/ui/input";
 import { Label } from "@/components/ui/label";
 import { Skeleton } from "@/components/ui/skeleton";
-import { ScrollArea } from "@/components/ui/scroll-area";
 import {
   Dialog,
   DialogContent,
@@ -673,11 +672,11 @@ function RunLogDialog({ run, onClose }: { run: AgentRun | null; onClose: () => v
         ) : error ? (
           <p className="text-sm text-destructive">שגיאה בטעינת הלוג: {String(error)}</p>
         ) : (
-          <ScrollArea className="h-[60vh] rounded-md border border-rule-soft bg-rule-soft/20 p-3">
+          <div className="h-[60vh] w-full overflow-y-auto overflow-x-hidden rounded-md border border-rule-soft bg-rule-soft/20 p-3">
-            <pre dir="ltr" className="text-[0.72rem] leading-relaxed whitespace-pre-wrap break-words text-navy text-start">
+            <pre dir="ltr" className="w-full min-w-0 max-w-full overflow-x-hidden text-[0.72rem] leading-relaxed whitespace-pre-wrap break-all text-navy text-start">
               {text || "אין פלט עדיין."}
             </pre>
-          </ScrollArea>
+          </div>
         )}
       </DialogContent>
     </Dialog>

web-ui/src/app/page.tsx

@@ -69,8 +69,10 @@ export default function HomePage() {
         {/* KPI row — mockup 04 .kpis (4-up, gold-washed "ממתינים לאישור") */}
         <KPICards cases={data} loading={isPending} />
 
-        {/* two-column body — main flow + narrow gold gate rail (mockup 04 .cols) */}
+        {/* two-column body — main flow + narrow gold gate rail (mockup 04 .cols).
-        <div className="grid gap-6 lg:grid-cols-[1fr_360px]">
+            Rail trimmed 360→280 so the cases table gets the width it needs and
+            no longer needs a horizontal scrollbar (chair request). */}
+        <div className="grid gap-6 lg:grid-cols-[1fr_280px]">
           <div className="space-y-6 min-w-0">
             {/* "תיקים לפי סטטוס" (פסים אופקיים) הוסר — פיזור-הסטטוסים מוצג
                 בדונאט "פיזור סטטוסים" בטור-הצד (#1). */}

web-ui/src/app/precedents/[id]/page.tsx

@@ -236,7 +236,11 @@ export default function PrecedentDetailPage({
 
           {/* side rail — citations + corroboration */}
           <div className="space-y-4">
-            <RelatedCasesSection caseId={id} related={data.related_cases ?? []} />
+            <RelatedCasesSection
+              caseId={id}
+              related={data.related_cases ?? []}
+              incoming={data.incoming_citations ?? []}
+            />
           </div>
         </div>
       </div>

web-ui/src/components/cases/agent-activity-feed.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useRef, useState, useEffect } from "react";
+import { useRef, useState, useEffect, useMemo } from "react";
 import { Button } from "@/components/ui/button";
 import { Textarea } from "@/components/ui/textarea";
 import { Badge } from "@/components/ui/badge";
@@ -151,7 +151,7 @@ function CommentCard({
               {identifier}
             </Badge>
           )}
-          <span className="text-[11px] text-ink-faint mr-auto flex items-center gap-1">
+          <span className="text-[11px] text-ink-faint me-auto flex items-center gap-1">
             <Clock className="w-3 h-3" />
             {timeAgo(comment.created_at)}
           </span>
@@ -267,7 +267,7 @@ function AskUserQuestionsForm({
           <fieldset key={q.id} className="space-y-2">
             <legend className="text-sm font-semibold text-navy mb-1">
               {q.prompt}
-              {(q.required ?? true) && <span className="text-rose-600 mr-1">*</span>}
+              {(q.required ?? true) && <span className="text-rose-600 ms-1">*</span>}
             </legend>
             <div className="space-y-1.5">
               {q.options.map((opt) => {
@@ -316,7 +316,7 @@ function AskUserQuestionsForm({
           {pending ? (
             <Loader2 className="w-4 h-4 animate-spin" />
           ) : (
-            <Send className="w-4 h-4 ml-1" />
+            <Send className="w-4 h-4 me-1" />
           )}
           {interaction.payload.submitLabel || "שלח תשובה"}
         </Button>
@@ -397,14 +397,14 @@ function RequestConfirmationForm({
           onClick={handleReject}
           disabled={pending || (requireReason && !reason.trim())}
         >
-          <XCircle className="w-4 h-4 ml-1" />
+          <XCircle className="w-4 h-4 me-1" />
           {rejectLabel}
         </Button>
         <Button size="sm" onClick={onAccept} disabled={pending}>
           {pending ? (
             <Loader2 className="w-4 h-4 animate-spin" />
           ) : (
-            <CheckCircle2 className="w-4 h-4 ml-1" />
+            <CheckCircle2 className="w-4 h-4 me-1" />
           )}
           {acceptLabel}
         </Button>
@@ -489,7 +489,7 @@ function SuggestTasksForm({
           onClick={() => (showReason ? onReject(reason.trim()) : setShowReason(true))}
           disabled={pending}
         >
-          <XCircle className="w-4 h-4 ml-1" />
+          <XCircle className="w-4 h-4 me-1" />
           {showReason ? "אישור דחייה" : "דחייה"}
         </Button>
         <Button
@@ -500,7 +500,7 @@ function SuggestTasksForm({
           {pending ? (
             <Loader2 className="w-4 h-4 animate-spin" />
           ) : (
-            <CheckCircle2 className="w-4 h-4 ml-1" />
+            <CheckCircle2 className="w-4 h-4 me-1" />
           )}
           אישור משימות נבחרות ({selected.size})
         </Button>
@@ -575,7 +575,7 @@ function InteractionCard({
               {identifier}
             </Badge>
           )}
-          <span className="text-[11px] text-ink-faint mr-auto flex items-center gap-1">
+          <span className="text-[11px] text-ink-faint me-auto flex items-center gap-1">
             <Clock className="w-3 h-3" />
             {timeAgo(interaction.resolved_at ?? interaction.created_at)}
           </span>
@@ -635,13 +635,13 @@ export function AgentActivityFeed({
   const [body, setBody] = useState("");
   const endRef = useRef<HTMLDivElement>(null);
 
-  // Build issue_id → identifier map
+  // Build issue_id → identifier map (memoized — the feed refetches every 10s,
-  const issueMap = new Map<string, string>();
+  // and a fresh Map each render would defeat the child cards' memoization).
-  if (data?.issues) {
+  const issueMap = useMemo(() => {
-    for (const iss of data.issues) {
+    const m = new Map<string, string>();
-      issueMap.set(iss.id, iss.identifier);
+    for (const iss of data?.issues ?? []) m.set(iss.id, iss.identifier);
-    }
+    return m;
-  }
+  }, [data?.issues]);
 
   // Auto-scroll on new comments or interactions
   const commentCount = data?.comments?.length ?? 0;
@@ -669,7 +669,7 @@ export function AgentActivityFeed({
   if (isLoading) {
     return (
       <div className="flex items-center justify-center py-12 text-ink-faint">
-        <Loader2 className="w-5 h-5 animate-spin ml-2" />
+        <Loader2 className="w-5 h-5 animate-spin me-2" />
         <span>טוען פעילות סוכנים...</span>
       </div>
     );
@@ -777,6 +777,7 @@ export function AgentActivityFeed({
           value={body}
           onChange={(e) => setBody(e.target.value)}
           placeholder="כתוב הוראה לסוכנים..."
+          aria-label="הוראה לסוכנים"
           className="min-h-[60px] resize-none text-sm"
           dir="rtl"
           onKeyDown={(e) => {
@@ -798,7 +799,7 @@ export function AgentActivityFeed({
             {sendComment.isPending ? (
               <Loader2 className="w-4 h-4 animate-spin" />
             ) : (
-              <Send className="w-4 h-4 ml-1" />
+              <Send className="w-4 h-4 me-1" />
             )}
             שלח
           </Button>

web-ui/src/components/cases/agent-status-widget.tsx

@@ -1,8 +1,20 @@
 "use client";
 
-import { useAgentActivity } from "@/lib/api/agents";
+import { useState } from "react";
+import { useAgentActivity, useResetCaseAgents } from "@/lib/api/agents";
 import type { PaperclipAgentStatus } from "@/lib/api/agents";
-import { Bot } from "lucide-react";
+import { Bot, RotateCcw, Loader2 } from "lucide-react";
+import { toast } from "sonner";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+  DialogTrigger,
+} from "@/components/ui/dialog";
+import { Button } from "@/components/ui/button";
 
 /* ── Status dot colors ───────────────────────────────────────── */
 
@@ -33,13 +45,82 @@ function AgentRow({ agent }: { agent: PaperclipAgentStatus }) {
         className={`w-2 h-2 rounded-full flex-shrink-0 ${statusDot(agent.status)}`}
       />
       <span className="text-xs text-ink truncate">{agent.name}</span>
-      <span className="text-[10px] text-ink-faint mr-auto">
+      <span className="text-[10px] text-ink-faint ms-auto">
         {STATUS_LABEL[agent.status] ?? agent.status}
       </span>
     </div>
   );
 }
 
+/* ── Reset button ─────────────────────────────────────────────── */
+
+function ResetAgentsButton({ caseNumber }: { caseNumber: string }) {
+  const [open, setOpen] = useState(false);
+  const reset = useResetCaseAgents(caseNumber);
+
+  function handleConfirm() {
+    reset.mutate(undefined, {
+      onSuccess: (result) => {
+        setOpen(false);
+        const agentNames = result.reset_agents
+          .filter((a) => a.ok)
+          .map((a) => a.name)
+          .join(", ");
+        const msg = agentNames
+          ? `סוכנים אופסו: ${agentNames}`
+          : "אופסו בהצלחה";
+        toast.success(msg);
+      },
+      onError: () => {
+        setOpen(false);
+        toast.error("שגיאה בביצוע האיפוס");
+      },
+    });
+  }
+
+  return (
+    <Dialog open={open} onOpenChange={setOpen}>
+      <DialogTrigger asChild>
+        <button
+          className="text-[10px] text-ink-faint hover:text-red-600 flex items-center gap-0.5 transition-colors"
+          title="אפס סוכנים תקועים"
+        >
+          <RotateCcw className="w-3 h-3" />
+          <span>אפס</span>
+        </button>
+      </DialogTrigger>
+      <DialogContent>
+        <DialogHeader>
+          <DialogTitle>איפוס סוכנים</DialogTitle>
+          <DialogDescription>
+            פעולה זו תאפס את כל הסוכנים במצב שגיאה ותחזיר issues פתוחים לניהול ידני.
+            פעולה הפיכה — ניתן להפעיל את הסוכנים מחדש אחרי האיפוס.
+          </DialogDescription>
+        </DialogHeader>
+        <DialogFooter>
+          <Button variant="outline" onClick={() => setOpen(false)} disabled={reset.isPending}>
+            ביטול
+          </Button>
+          <Button
+            variant="destructive"
+            onClick={handleConfirm}
+            disabled={reset.isPending}
+          >
+            {reset.isPending ? (
+              <>
+                <Loader2 className="w-3.5 h-3.5 ms-1.5 animate-spin" />
+                מאפס...
+              </>
+            ) : (
+              "אפס סוכנים"
+            )}
+          </Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}
+
 /* ── Widget ───────────────────────────────────────────────────── */
 
 export function AgentStatusWidget({
@@ -54,6 +135,7 @@ export function AgentStatusWidget({
 
   const agents = data.agents ?? [];
   const activeCount = agents.filter((a) => a.status === "active" || a.status === "running").length;
+  const hasErrors = agents.some((a) => a.status === "error");
 
   return (
     <div className="space-y-2">
@@ -62,11 +144,14 @@ export function AgentStatusWidget({
           <Bot className="w-3.5 h-3.5" />
           <span>סוכנים</span>
         </div>
-        {agents.length > 0 && (
+        <div className="flex items-center gap-2">
-          <span className="text-[10px] text-ink-faint">
+          {hasErrors && <ResetAgentsButton caseNumber={caseNumber} />}
-            {activeCount} פעילים מתוך {agents.length}
+          {agents.length > 0 && (
-          </span>
+            <span className="text-[10px] text-ink-faint">
-        )}
+              {activeCount} פעילים מתוך {agents.length}
+            </span>
+          )}
+        </div>
       </div>
 
       <div className="space-y-0.5">

web-ui/src/components/cases/case-edit-dialog.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useEffect, useState } from "react";
+import { useState } from "react";
 import { useForm, Controller } from "react-hook-form";
 import { zodResolver } from "@hookform/resolvers/zod";
 import { toast } from "sonner";
@@ -54,22 +54,26 @@ export function CaseEditDialog({ data }: { data: CaseDetail }) {
     },
   });
 
-  /* Re-sync the form when the underlying case refetches after save */
+  /* Reset to the latest case values only on the open→true transition.
-  useEffect(() => {
+   * Resetting on every `data` change would clobber in-progress edits, because
-    if (!open) return;
+   * useCase refetches every 5s (refetchInterval) while the dialog is open. */
-    form.reset({
+  const handleOpenChange = (next: boolean) => {
-      title: data.title ?? "",
+    if (next) {
-      subject: data.subject ?? "",
+      form.reset({
-      hearing_date: data.hearing_date ?? "",
+        title: data.title ?? "",
-      notes: "",
+        subject: data.subject ?? "",
-      expected_outcome: data.expected_outcome ?? "",
+        hearing_date: data.hearing_date ?? "",
-      appellants: data.appellants ?? [],
+        notes: "",
-      respondents: data.respondents ?? [],
+        expected_outcome: data.expected_outcome ?? "",
-      property_address: data.property_address ?? "",
+        appellants: data.appellants ?? [],
-      permit_number: data.permit_number ?? "",
+        respondents: data.respondents ?? [],
-      proceeding_type: data.proceeding_type ?? "ערר",
+        property_address: data.property_address ?? "",
-    });
+        permit_number: data.permit_number ?? "",
-  }, [open, data, form]);
+        proceeding_type: data.proceeding_type ?? "ערר",
+      });
+    }
+    setOpen(next);
+  };
 
   const onSubmit = form.handleSubmit(async (values) => {
     try {
@@ -82,7 +86,7 @@ export function CaseEditDialog({ data }: { data: CaseDetail }) {
   });
 
   return (
-    <Dialog open={open} onOpenChange={setOpen}>
+    <Dialog open={open} onOpenChange={handleOpenChange}>
       <DialogTrigger asChild>
         <Button variant="outline" size="sm">
           עריכת פרטי תיק

web-ui/src/components/cases/case-header.tsx

@@ -63,8 +63,10 @@ export function CaseHeader({
         <span className="text-navy tabular-nums">{data?.case_number ?? "…"}</span>
       </nav>
 
-      <div className="flex items-start justify-between gap-6 flex-wrap">
+      {/* title block + metadata on one row — no wrap, so the metadata dl stays
-        <div className="min-w-0">
+          parallel to the H1 and a long parties line can't push it below. */}
+      <div className="flex items-start justify-between gap-6">
+        <div className="min-w-0 flex-1">
           {/* title row — H1 + status/type/blam chips inline (mockup .band h1) */}
           <h1 className="text-navy text-[1.7rem] font-bold leading-tight flex items-center gap-3 flex-wrap mb-0">
             <span className="tabular-nums">
@@ -107,11 +109,15 @@ export function CaseHeader({
               {data.title}
             </p>
           )}
-          {/* parties line (mockup .parties) */}
+          {/* parties line (mockup .parties) — clamped to 2 lines so a long list
+              of appellants/respondents doesn't grow the band; full text on hover. */}
           {parties ? (
-            <p className="text-ink-soft text-sm mt-1.5">{parties}</p>
+            <p className="text-ink-soft text-sm mt-1.5 max-w-3xl line-clamp-2" title={parties}>
+              {parties}
+            </p>
           ) : data?.subject ? (
-            <p className="text-ink-soft text-sm mt-1.5 max-w-3xl leading-relaxed">
+            <p className="text-ink-soft text-sm mt-1.5 max-w-3xl leading-relaxed line-clamp-2"
+               title={data.subject}>
               {data.subject}
             </p>
           ) : null}

web-ui/src/components/cases/cases-table.tsx

@@ -120,7 +120,7 @@ const columns: ColumnDef<Case>[] = [
     accessorKey: "title",
     header: "כותרת",
     cell: ({ row }) => (
-      <div className="text-ink max-w-[420px] truncate flex items-center gap-2" title={row.original.title}>
+      <div className="text-ink max-w-[420px] min-w-0 truncate flex items-center gap-2" title={row.original.title}>
         {(row.original.proceeding_type === 'בל"מ' || isBlamSubtype(row.original.appeal_subtype)) && (
           <Badge
             variant="outline"

web-ui/src/components/cases/decision-blocks-panel.tsx

@@ -56,7 +56,7 @@ export function DecisionBlocksPanel({ caseNumber }: { caseNumber: string }) {
   }
   if (!data) return null;
 
-  const written = data.blocks.filter((b) => b.word_count > 0).length;
+  const written = data.blocks.filter((b) => (b.word_count ?? 0) > 0).length;
 
   return (
     <div className="space-y-4">
@@ -101,11 +101,11 @@ export function DecisionBlocksPanel({ caseNumber }: { caseNumber: string }) {
                   {blockLabel(block)}
                 </span>
                 <Badge
-                  className={`text-[0.65rem] border ${STATUS_CLASSES[block.status]}`}
+                  className={`text-[0.65rem] border ${STATUS_CLASSES[block.status] ?? ""}`}
                 >
-                  {STATUS_LABELS[block.status]}
+                  {STATUS_LABELS[block.status] ?? block.status}
                 </Badge>
-                {block.word_count > 0 && (
+                {(block.word_count ?? 0) > 0 && (
                   <span className="text-[0.7rem] text-ink-muted tabular-nums">
                     {block.word_count} מילים
                   </span>
@@ -137,19 +137,22 @@ function BlockEditor({
   caseNumber: string;
   block: DecisionBlock;
 }) {
+  // The endpoint has no FastAPI response model, so `content` may arrive null
+  // for an empty block — coerce to "" to keep `.trim()` / state safe.
+  const content = block.content ?? "";
   const [editing, setEditing] = useState(false);
-  const [value, setValue] = useState(block.content);
+  const [value, setValue] = useState(content);
   const [state, setState] = useState<SaveState>({ kind: "idle" });
   /* The last content known to be persisted — used to skip no-op saves. */
-  const [baseline, setBaseline] = useState(block.content);
+  const [baseline, setBaseline] = useState(content);
   const save = useSaveBlock(caseNumber);
 
   /* Re-sync when the upstream query refetches (e.g. after another save) while
    * not actively editing. Adjusting state during render — the documented React
    * pattern for derived-from-props — avoids a setState-in-effect cascade. */
-  if (!editing && block.content !== baseline) {
+  if (!editing && content !== baseline) {
-    setBaseline(block.content);
+    setBaseline(content);
-    setValue(block.content);
+    setValue(content);
   }
 
   async function handleSave() {
@@ -171,8 +174,8 @@ function BlockEditor({
   if (!editing) {
     return (
       <div className="space-y-3">
-        {block.content.trim() ? (
+        {content.trim() ? (
-          <Markdown content={block.content} />
+          <Markdown content={content} />
         ) : (
           <p className="text-sm text-ink-muted italic">בלוק ריק.</p>
         )}

web-ui/src/components/cases/document-type-editor.tsx

@@ -90,13 +90,23 @@ export function DocumentTypeEditor({
     // clear it so it doesn't dangle confusingly in metadata.
     if (!isAppraisal && appraiserSide) body.appraiser_side = "";
 
-    await patch.mutateAsync({ docId, patch: body });
+    // Swallow the rejection — errors surface via `patch.isError`; an unhandled
-    setSaved(true);
+    // rejection from this async click handler would otherwise leak.
+    try {
+      await patch.mutateAsync({ docId, patch: body });
+      setSaved(true);
+    } catch {
+      /* surfaced via patch.isError */
+    }
   }
 
   async function handleExtract() {
-    const result = await extract.mutateAsync();
+    try {
-    setExtractResult(result);
+      const result = await extract.mutateAsync();
+      setExtractResult(result);
+    } catch {
+      /* surfaced via extract.isError */
+    }
   }
 
   // Build the on-badge label: "שומה · שמאי הוועדה" when both present.
@@ -301,7 +311,7 @@ function PostSaveView({
             נותרו {extractResult.missing.length} שומות ללא תיוג צד. תייג אותן
             לפני הפעלת החילוץ:
           </p>
-          <ul className="list-disc pr-4 space-y-0.5">
+          <ul className="list-disc ps-4 space-y-0.5">
             {extractResult.missing.map((m) => (
               <li key={m.document_id} className="truncate">
                 {m.title}

web-ui/src/components/cases/drafts-panel.tsx

@@ -1,6 +1,7 @@
 "use client";
 
 import { useRef, useState } from "react";
+import Link from "next/link";
 import { useQueryClient } from "@tanstack/react-query";
 import { Badge } from "@/components/ui/badge";
 import { Button } from "@/components/ui/button";
@@ -344,9 +345,9 @@ export function DraftsPanel({
             </Button>
             <span className="text-[0.7rem] text-ink-muted">
               סטטוס הריצה — בדף{" "}
-              <a href="/operations" className="underline">
+              <Link href="/operations" className="underline">
                 התפעול
-              </a>
+              </Link>
             </span>
           </div>
         )}
@@ -628,7 +629,7 @@ function CitationsSection({ caseNumber }: { caseNumber: string }) {
 
       <div className="rounded-lg border border-rule overflow-hidden divide-y divide-rule">
         {data.linked.map((c) => (
-          <a
+          <Link
             key={`l-${c.citation}`}
             href={`/precedents/${c.cited_id}`}
             className="flex items-center gap-2 px-4 py-2.5 text-sm hover:bg-rule-soft/20"
@@ -641,7 +642,7 @@ function CitationsSection({ caseNumber }: { caseNumber: string }) {
             <Badge className="ms-auto bg-success-bg text-success border-success/40 text-[0.65rem] shrink-0">
               בספרייה
             </Badge>
-          </a>
+          </Link>
         ))}
         {data.missing.map((c) => (
           <div

web-ui/src/components/cases/legal-arguments-panel.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useMemo } from "react";
+import { useMemo, type ReactNode } from "react";
 import {
   Accordion,
   AccordionContent,
@@ -9,7 +9,6 @@ import {
 } from "@/components/ui/accordion";
 import { Badge } from "@/components/ui/badge";
 import { Button } from "@/components/ui/button";
-import { Card, CardContent } from "@/components/ui/card";
 import {
   Popover,
   PopoverContent,
@@ -22,6 +21,7 @@ import {
   PRIORITY_ORDER,
   useAggregateArguments,
   useLegalArguments,
+  type AggregateArgumentsResult,
   type LegalArgument,
   type LegalArgumentParty,
   type LegalArgumentPriority,
@@ -189,82 +189,167 @@ export function LegalArgumentsPanel({ caseNumber }: LegalArgumentsPanelProps) {
   }, [data]);
 
   const handleAggregate = (force: boolean) => {
+    // Status feedback is rendered inline as a banner from `aggregate.data`;
+    // only hard transport/HTTP errors fall through to a toast.
     aggregate.mutate(force, {
-      onSuccess: () => {
-        toast.success(
-          force
-            ? "הופעלה חזרה חישוב טיעונים (force). יסתיים תוך דקה."
-            : "הופעל חישוב טיעונים. רענן בעוד דקה.",
-        );
-      },
       onError: (e) => toast.error(`שגיאה: ${(e as Error).message}`),
     });
   };
 
   return (
-    <Card className="bg-surface border-rule shadow-sm">
+    <div className="space-y-4">
-      <CardContent className="px-6 py-5 space-y-4">
+      <div className="flex items-center justify-between flex-wrap gap-3">
-        <div className="flex items-center justify-between flex-wrap gap-3">
+        <div>
-          <div>
+          <h2 className="text-navy text-base font-semibold">טיעונים משפטיים</h2>
-            <h2 className="text-navy text-base font-semibold">
+          <p className="text-ink-muted text-xs mt-0.5">
-              טיעונים משפטיים
+            טיעונים מאוגדים מתוך הטענות הגולמיות, מקובצים לפי צד וקדימות. החישוב
-            </h2>
+            רץ אצל המנתח המשפטי ומדווח בחזרה.
-            <p className="text-ink-muted text-xs mt-0.5">
+          </p>
-              טיעונים מאוגדים מתוך הפרופוזיציות הגולמיות, מקובצים לפי צד וקדימות.
-            </p>
-          </div>
-          <div className="flex items-center gap-2">
-            <Button
-              variant="outline"
-              size="sm"
-              disabled={aggregate.isPending}
-              onClick={() => handleAggregate(false)}
-            >
-              {aggregate.isPending ? (
-                <Loader2 className="w-3.5 h-3.5 animate-spin me-1.5" />
-              ) : (
-                <Sparkles className="w-3.5 h-3.5 me-1.5" />
-              )}
-              חשב טיעונים
-            </Button>
-            <Button
-              variant="ghost"
-              size="sm"
-              disabled={aggregate.isPending || !data?.total}
-              onClick={() => handleAggregate(true)}
-              title="חישוב מחדש (מוחק טיעונים קיימים)"
-            >
-              <RefreshCw className="w-3.5 h-3.5" />
-            </Button>
-          </div>
         </div>
+        <div className="flex items-center gap-2">
+          <Button
+            variant="outline"
+            size="sm"
+            disabled={aggregate.isPending}
+            onClick={() => handleAggregate(false)}
+          >
+            {aggregate.isPending ? (
+              <Loader2 className="w-3.5 h-3.5 animate-spin me-1.5" />
+            ) : (
+              <Sparkles className="w-3.5 h-3.5 me-1.5" />
+            )}
+            חשב טיעונים
+          </Button>
+          <Button
+            variant="ghost"
+            size="sm"
+            disabled={aggregate.isPending || !data?.total}
+            onClick={() => handleAggregate(true)}
+            title="חישוב מחדש (מוחק טיעונים קיימים)"
+          >
+            <RefreshCw className="w-3.5 h-3.5" />
+          </Button>
+        </div>
+      </div>
 
-        {isPending ? (
+      <AggregateStatusBanner
-          <div className="space-y-2">
+        result={aggregate.data}
-            <Skeleton className="h-6 w-48" />
+        force={aggregate.variables ?? false}
-            <Skeleton className="h-20 w-full" />
+      />
-            <Skeleton className="h-20 w-full" />
+
-          </div>
+      {isPending ? (
-        ) : isError ? (
+        <div className="space-y-2">
-          <p className="text-danger text-sm">
+          <Skeleton className="h-6 w-48" />
-            שגיאה בטעינת טיעונים: {(error as Error).message}
+          <Skeleton className="h-20 w-full" />
-          </p>
+          <Skeleton className="h-20 w-full" />
-        ) : !data?.total ? (
+        </div>
-          <p className="text-ink-muted text-sm">
+      ) : isError ? (
-            אין טיעונים מאוגדים עדיין. לחץ &ldquo;חשב טיעונים&rdquo; כדי להריץ את ה-aggregator.
+        <p className="text-danger text-sm">
-          </p>
+          שגיאה בטעינת טיעונים: {(error as Error).message}
-        ) : (
+        </p>
-          <div className="space-y-6">
+      ) : !data?.total ? (
-            {parties.map((party) => (
+        <p className="text-ink-muted text-sm">
-              <PartySection
+          אין טיעונים מאוגדים עדיין. לחץ &ldquo;חשב טיעונים&rdquo; כדי לשלוח את
-                key={party}
+          החישוב למנתח המשפטי.
-                party={party}
+        </p>
-                args={data.by_party[party] ?? []}
+      ) : (
-              />
+        <div className="space-y-6">
-            ))}
+          {parties.map((party) => (
-          </div>
+            <PartySection
-        )}
+              key={party}
-      </CardContent>
+              party={party}
-    </Card>
+              args={data.by_party[party] ?? []}
+            />
+          ))}
+        </div>
+      )}
+    </div>
+  );
+}
+
+const BANNER_TONE = {
+  info: "border-info/30 bg-info-bg",
+  gold: "border-gold/40 bg-gold-wash",
+  muted: "border-rule bg-rule-soft",
+  warn: "border-warn/40 bg-warn-bg",
+} as const;
+
+const DOT_TONE = {
+  info: "bg-info",
+  gold: "bg-gold-deep",
+  muted: "bg-ink-light",
+  warn: "bg-warn",
+} as const;
+
+const TITLE_TONE = {
+  info: "text-info",
+  gold: "text-gold-deep",
+  muted: "text-ink-soft",
+  warn: "text-warn",
+} as const;
+
+/**
+ * Inline status feedback for the aggregation trigger — mirrors the approved
+ * Claude Design mockup (25-legal-arguments-panel). Shown only after a click;
+ * the four states map to the endpoint's discriminated-union result.
+ */
+function AggregateStatusBanner({
+  result,
+  force,
+}: {
+  result: AggregateArgumentsResult | undefined;
+  force: boolean;
+}) {
+  if (!result) return null;
+
+  let tone: keyof typeof BANNER_TONE;
+  let title: string;
+  let body: ReactNode;
+
+  switch (result.status) {
+    case "queued":
+      tone = "info";
+      title = "נשלח לאנליטיקאי.";
+      body = `${force ? "החישוב-מחדש" : "החישוב"} רץ ברקע אצל המנתח המשפטי; התוצאה תופיע תוך כמה דקות — רענן את הדף.`;
+      break;
+    case "exists":
+      tone = "gold";
+      title = "כבר חושב.";
+      body = result.message;
+      break;
+    case "no_claims":
+      tone = "muted";
+      title = "אין טענות גולמיות.";
+      body = result.message;
+      break;
+    case "skipped":
+      tone = "warn";
+      title = "לא ניתן להפעיל אוטומטית";
+      body = (
+        <>
+          {` (${result.reason}). הרץ ידנית מ-Claude Code: `}
+          <code className="font-mono text-[0.72rem] bg-navy/5 rounded px-1.5 py-0.5 select-all">
+            mcp__legal-ai__aggregate_claims_to_arguments
+          </code>
+        </>
+      );
+      break;
+    default:
+      return null;
+  }
+
+  return (
+    <div
+      className={`flex items-start gap-2.5 rounded-md border px-3.5 py-2.5 text-[0.8rem] leading-relaxed text-ink-soft ${BANNER_TONE[tone]}`}
+    >
+      <span
+        className={`mt-1.5 size-2 flex-none rounded-full ${DOT_TONE[tone]}`}
+        aria-hidden
+      />
+      <p>
+        <strong className={`font-semibold ${TITLE_TONE[tone]}`}>{title}</strong>{" "}
+        {body}
+      </p>
+    </div>
   );
 }

web-ui/src/components/cases/status-changer.tsx

@@ -25,14 +25,20 @@ export function StatusChanger({
   caseNumber: string;
   currentStatus?: CaseStatus;
 }) {
-  const [selected, setSelected] = useState<CaseStatus | "">(currentStatus ?? "");
+  // `null` = untouched → the dropdown tracks the live `currentStatus` (which
+  // arrives async and changes on the 5s poll / external updates). Only an
+  // explicit pick overrides it, until save resets back to tracking.
+  const [picked, setPicked] = useState<CaseStatus | null>(null);
   const mutate = useUpdateCase(caseNumber);
+  const effective: CaseStatus | "" = picked ?? currentStatus ?? "";
+  const canSave = Boolean(effective) && effective !== currentStatus;
 
   const handleSave = async () => {
-    if (!selected || selected === currentStatus) return;
+    if (!canSave || !effective) return;
     try {
-      await mutate.mutateAsync({ status: selected });
+      await mutate.mutateAsync({ status: effective });
-      toast.success(`הסטטוס עודכן ל${STATUS_LABELS[selected]}`);
+      toast.success(`הסטטוס עודכן ל${STATUS_LABELS[effective]}`);
+      setPicked(null);
     } catch (e) {
       toast.error(e instanceof Error ? e.message : "שגיאה בעדכון הסטטוס");
     }
@@ -43,8 +49,8 @@ export function StatusChanger({
       <label className="text-[0.72rem] text-ink-muted block">שינוי סטטוס ידני</label>
       <div className="flex items-center gap-2">
         <Select
-          value={selected || "__current__"}
+          value={effective || "__current__"}
-          onValueChange={(v) => setSelected(v === "__current__" ? "" : v as CaseStatus)}
+          onValueChange={(v) => setPicked(v === "__current__" ? null : v as CaseStatus)}
           dir="rtl"
         >
           <SelectTrigger className="text-[0.75rem] h-8">
@@ -68,7 +74,7 @@ export function StatusChanger({
           size="sm"
           variant="outline"
           className="h-8 text-[0.72rem] px-3 shrink-0"
-          disabled={!selected || selected === currentStatus || mutate.isPending}
+          disabled={!canSave || mutate.isPending}
           onClick={handleSave}
         >
           {mutate.isPending ? "שומר…" : "עדכן"}

web-ui/src/components/compose/citation-verification-panel.tsx

@@ -0,0 +1,261 @@
+"use client";
+
+/**
+ * "אימות פסיקה" tab (X11 Phase 2 / #154) — per legal argument, the in-corpus
+ * supporting precedents with the cumulative authority signal (cited_by), a verify
+ * gate + chair note, and per-issue radar (unlinked digests). The writer cites only
+ * verified quotes (INV-AH); the digest is never cited (INV-DIG1, radar only).
+ */
+
+import { useState } from "react";
+import { toast } from "sonner";
+import { Card, CardContent } from "@/components/ui/card";
+import { Skeleton } from "@/components/ui/skeleton";
+import {
+  useCitationVerification,
+  useVerifyCitation,
+  type ArgumentBlock,
+  type SupportingPrecedent,
+  type RadarLead,
+} from "@/lib/api/citation-verification";
+
+const PRIORITY_LABEL: Record<string, string> = {
+  threshold: "סף",
+  substantive: "מהותית",
+  procedural: "פרוצדורלית",
+  relief: "סעד",
+};
+
+const RADAR_LABEL: Record<string, string> = {
+  new_lead: "ליד חדש",
+  gap_open: "בתור-חסרים",
+  fetched: "נמשך",
+  available_link: "בקורפוס — לקשר",
+};
+
+export function CitationVerificationPanel({ caseNumber }: { caseNumber: string }) {
+  const { data, isPending, isError } = useCitationVerification(caseNumber);
+  const verify = useVerifyCitation(caseNumber);
+  const [notes, setNotes] = useState<Record<string, string>>({});
+
+  if (isPending) {
+    return (
+      <Card className="bg-surface border-rule shadow-sm">
+        <CardContent className="px-6 py-5 space-y-3">
+          <Skeleton className="h-6 w-64" />
+          <Skeleton className="h-24 w-full" />
+          <Skeleton className="h-24 w-full" />
+        </CardContent>
+      </Card>
+    );
+  }
+  if (isError || !data || data.status !== "ok") {
+    return (
+      <Card className="bg-surface border-rule shadow-sm">
+        <CardContent className="px-6 py-8 text-center text-ink-muted">
+          לא ניתן לטעון את אימות-הפסיקה כעת.
+        </CardContent>
+      </Card>
+    );
+  }
+  if (!data.arguments.length) {
+    return (
+      <Card className="bg-surface border-rule shadow-sm">
+        <CardContent className="px-6 py-12 text-center space-y-2">
+          <div className="text-gold text-3xl" aria-hidden>❦</div>
+          <p className="text-navy font-semibold mb-0">אין עדיין סוגיות מזוקקות לתיק</p>
+          <p className="text-ink-muted text-sm">
+            הרץ את ניתוח-הטענות (aggregate) כדי שהמערכת תזהה סוגיות ותתאים להן פסיקה.
+          </p>
+        </CardContent>
+      </Card>
+    );
+  }
+
+  function runVerify(a: ArgumentBlock, s: SupportingPrecedent, verified: boolean) {
+    verify.mutate(
+      {
+        argument_id: a.argument_id,
+        case_law_id: s.case_law_id,
+        quote: s.quote,
+        citation: s.case_number || s.case_name,
+        chair_note: notes[s.case_law_id] ?? s.chair_note,
+        verified,
+        attached_id: s.attached_id ?? "",
+      },
+      {
+        onSuccess: () => toast.success(verified ? "הציטוט אומת" : "סומן כלא-רלוונטי"),
+        onError: (e: Error) => toast.error(`שגיאה: ${e.message}`),
+      },
+    );
+  }
+
+  function saveNote(a: ArgumentBlock, s: SupportingPrecedent) {
+    const note = notes[s.case_law_id];
+    if (note === undefined || note === s.chair_note) return;
+    verify.mutate(
+      {
+        argument_id: a.argument_id,
+        case_law_id: s.case_law_id,
+        quote: s.quote,
+        citation: s.case_number || s.case_name,
+        chair_note: note,
+        verified: s.verified,
+        attached_id: s.attached_id ?? "",
+      },
+      {
+        onSuccess: () => toast.success("הערת-יו״ר נשמרה"),
+        onError: (e: Error) => toast.error(`שגיאה: ${e.message}`),
+      },
+    );
+  }
+
+  const sum = data.summary;
+
+  return (
+    <div className="space-y-4">
+      {/* summary */}
+      <div className="flex flex-wrap items-center gap-x-5 gap-y-1 text-sm text-ink-soft bg-parchment border border-rule rounded-lg px-4 py-2.5">
+        <span>סוגיות עם פסיקה: <b className="text-navy tabular-nums">{sum.arguments_with_support}/{sum.arguments_total}</b></span>
+        <span>אומתו: <b className="text-success tabular-nums">{sum.verified}</b></span>
+        <span>לידֵי רדאר: <b className="text-warn tabular-nums">{sum.radar_leads}</b></span>
+        <span className="text-ink-muted text-[0.78rem] ms-auto">ה-writer מצטט רק מאומתים (INV-AH) · היומון אינו מצוטט (INV-DIG1)</span>
+      </div>
+
+      {data.arguments.map((a) => (
+        <Card key={a.argument_id} className="bg-surface border-rule shadow-sm overflow-hidden">
+          <CardContent className="p-0">
+            {/* issue header */}
+            <div className="flex items-start gap-3 px-5 py-3.5 border-b border-rule-soft bg-parchment/60">
+              <div className="min-w-0 flex-1">
+                <div className="text-navy font-bold text-[0.98rem]">{a.title}</div>
+                <div className="text-ink-muted text-xs mt-0.5">
+                  סוגיה{a.legal_topic ? ` · ${a.legal_topic}` : ""}
+                </div>
+              </div>
+              {a.priority && (
+                <span className="shrink-0 rounded-full bg-gold-wash text-gold-deep text-[0.68rem] font-semibold px-2.5 py-0.5">
+                  {PRIORITY_LABEL[a.priority] ?? a.priority}
+                </span>
+              )}
+            </div>
+
+            {/* supporting precedents */}
+            <div className="px-5 py-3">
+              <div className="text-[0.72rem] font-bold text-ink-muted mb-1">פסיקה תומכת בקורפוס</div>
+              {a.supporting.length === 0 ? (
+                <p className="text-ink-muted text-sm py-2">לא נמצאה פסיקה תומכת בקורפוס לסוגיה זו.</p>
+              ) : (
+                <ul className="list-none p-0 m-0 divide-y divide-rule-soft">
+                  {a.supporting.map((s) => (
+                    <li key={s.case_law_id} className="py-3 first:pt-1">
+                      <div className="flex items-center gap-2 flex-wrap">
+                        <a
+                          href={`/precedents/${s.case_law_id}`}
+                          className="font-bold text-gold-deep text-[0.86rem] hover:text-navy"
+                        >
+                          {s.case_number || s.case_name}
+                        </a>
+                        {s.cited_by.positive > 0 && (
+                          <span className="rounded-full bg-success-bg text-success text-[0.68rem] font-semibold px-2 py-0.5">
+                            אומץ ×{s.cited_by.positive}
+                          </span>
+                        )}
+                        {s.cited_by.negative > 0 && (
+                          <span className="rounded-full bg-danger-bg text-danger text-[0.68rem] font-semibold px-2 py-0.5">
+                            אובחן ×{s.cited_by.negative}
+                          </span>
+                        )}
+                      </div>
+                      {s.quote && (
+                        <blockquote className="border-s-[3px] border-gold bg-gold-wash text-ink-soft text-sm leading-7 rounded-e-md px-3.5 py-2 my-2 mx-0">
+                          {s.quote}
+                        </blockquote>
+                      )}
+                      <div className="flex items-center gap-2 flex-wrap">
+                        {s.verified ? (
+                          <>
+                            <span className="text-success text-xs font-semibold">✓ אומת ע״י היו״ר</span>
+                            <button
+                              type="button"
+                              disabled={verify.isPending}
+                              onClick={() => runVerify(a, s, false)}
+                              className="text-xs text-ink-muted hover:text-danger border border-rule rounded-md px-2.5 py-1 disabled:opacity-40"
+                            >
+                              בטל אימות
+                            </button>
+                          </>
+                        ) : (
+                          <>
+                            <button
+                              type="button"
+                              disabled={verify.isPending}
+                              onClick={() => runVerify(a, s, true)}
+                              className="text-xs font-semibold text-success bg-success-bg border border-success/30 rounded-md px-3 py-1 hover:brightness-95 disabled:opacity-40"
+                            >
+                              ✓ מאמת
+                            </button>
+                            <button
+                              type="button"
+                              disabled={verify.isPending}
+                              onClick={() => runVerify(a, s, false)}
+                              className="text-xs font-semibold text-danger border border-rule rounded-md px-3 py-1 hover:bg-danger-bg disabled:opacity-40"
+                            >
+                              ✗ לא רלוונטי
+                            </button>
+                          </>
+                        )}
+                        {s.cited_by.negative > 0 && !s.verified && (
+                          <span className="text-warn text-[0.72rem] ms-1">⚠ אובחן — בדוק הקשר</span>
+                        )}
+                      </div>
+                      {/* chair note */}
+                      <div className="mt-2 flex items-start gap-2">
+                        <label className="text-[0.72rem] font-bold text-gold-deep whitespace-nowrap pt-1.5">
+                          הערת יו״ר:
+                        </label>
+                        <input
+                          type="text"
+                          defaultValue={s.chair_note}
+                          placeholder="הוסף הערה לציטוט — מדוע תומך / הסתייגות / איך לשלב…"
+                          onChange={(e) => setNotes((n) => ({ ...n, [s.case_law_id]: e.target.value }))}
+                          onBlur={() => saveNote(a, s)}
+                          className="flex-1 text-[0.8rem] text-ink-soft bg-parchment border border-dashed border-rule rounded-md px-2.5 py-1.5 focus:outline-none focus:border-gold"
+                        />
+                      </div>
+                    </li>
+                  ))}
+                </ul>
+              )}
+            </div>
+
+            {/* radar — unlinked digests for this issue (INV-DIG1: pointer, not citation) */}
+            <RadarStrip radar={a.radar} />
+          </CardContent>
+        </Card>
+      ))}
+    </div>
+  );
+}
+
+function RadarStrip({ radar }: { radar: RadarLead[] }) {
+  if (!radar.length) return null;
+  return (
+    <div className="mx-5 mb-4 rounded-lg border border-dashed border-gold bg-parchment px-4 py-2.5">
+      <div className="text-[0.72rem] font-bold text-gold-deep mb-1">
+        📡 רדאר — פסיקה שאין לנו בקורפוס (מצביע, לא מצוטט)
+      </div>
+      <ul className="list-none p-0 m-0 space-y-1">
+        {radar.map((r) => (
+          <li key={r.digest_id} className="flex items-center gap-2 flex-wrap text-[0.8rem]">
+            <span className="font-semibold text-navy">{r.underlying_citation || "(אין מראה-מקום)"}</span>
+            <span className="rounded-full bg-warn-bg text-warn text-[0.64rem] font-semibold px-2 py-0.5">
+              {RADAR_LABEL[r.action] ?? r.action}
+            </span>
+            <span className="text-ink-muted truncate">{r.headline}</span>
+          </li>
+        ))}
+      </ul>
+    </div>
+  );
+}

web-ui/src/components/documents/upload-sheet.tsx

@@ -43,7 +43,9 @@ function statusLabel(event: ProgressEvent | null): string {
   if (event.status === "processing")
     return event.step ? `בעיבוד · ${event.step}` : "בעיבוד";
   if (event.status === "completed") return "הושלם";
-  if (event.status === "unknown") return "הושלם";
+  // TTL expired / subscribed too late — the outcome is genuinely unknown, not
+  // a confirmed success. The case-detail refetch is the real source of truth.
+  if (event.status === "unknown") return "הסתיים — רענן לאישור";
   if (event.status === "failed") return event.error ?? "נכשל";
   return event.status;
 }
@@ -62,7 +64,9 @@ function UploadRowView({ row, caseNumber }: { row: UploadRow; caseNumber: string
   const progress = useProgress(row.taskId, caseNumber);
   const pct = row.error ? 100 : progressPercent(progress);
   const failed = row.error || progress?.status === "failed";
-  const done = progress?.status === "completed" || progress?.status === "unknown";
+  const done = progress?.status === "completed";
+  // `unknown` = settled-but-indeterminate; render neutral, not green success.
+  const indeterminate = progress?.status === "unknown";
 
   return (
     <li className="rounded-lg border border-rule bg-parchment/40 px-4 py-3 space-y-2">
@@ -71,6 +75,8 @@ function UploadRowView({ row, caseNumber }: { row: UploadRow; caseNumber: string
           <CheckCircle2 className="w-4 h-4 text-success shrink-0" />
         ) : failed ? (
           <XCircle className="w-4 h-4 text-danger shrink-0" />
+        ) : indeterminate ? (
+          <CheckCircle2 className="w-4 h-4 text-ink-muted shrink-0" />
         ) : (
           <Loader2 className="w-4 h-4 text-gold animate-spin shrink-0" />
         )}

web-ui/src/components/graph/graph-filter-panel.tsx

@@ -48,7 +48,14 @@ export type GraphControls = {
 };
 
 const ALL = "__all__";
-const YEARS = Array.from({ length: 2026 - 1994 + 1 }, (_, i) => 2026 - i);
+// Floor covers the oldest dated precedent in the corpus (currently ע"א 725/81 →
+// 1982); ceiling tracks the current year so the range never ages out.
+const YEAR_FLOOR = 1980;
+const CURRENT_YEAR = new Date().getFullYear();
+const YEARS = Array.from(
+  { length: CURRENT_YEAR - YEAR_FLOOR + 1 },
+  (_, i) => CURRENT_YEAR - i,
+);
 
 const COLOR_BY: { value: ColorBy; label: string }[] = [
   { value: "type", label: "סוג נקודה" },

web-ui/src/components/missing-precedents/missing-precedents-table.tsx

@@ -1,7 +1,7 @@
 "use client";
 
-import { useState } from "react";
+import { useMemo, useState } from "react";
-import { Trash2, Upload, Pencil, ExternalLink } from "lucide-react";
+import { Trash2, Upload, Pencil, ExternalLink, ChevronDown } from "lucide-react";
 import { toast } from "sonner";
 import Link from "next/link";
 import {
@@ -63,6 +63,24 @@ function SourceChip({ party }: { party: CitedByParty | null }) {
   );
 }
 
+const COLS = 7;
+
+function TableHeaderRow() {
+  return (
+    <TableHeader className="bg-parchment">
+      <TableRow className="border-rule hover:bg-transparent">
+        <TableHead className="text-ink-muted text-right font-medium text-xs">פסיקה</TableHead>
+        <TableHead className="text-ink-muted text-right font-medium text-xs">נושא</TableHead>
+        <TableHead className="text-ink-muted text-right font-medium text-xs">תיק</TableHead>
+        <TableHead className="text-ink-muted text-right font-medium text-xs">צוטט ע״י</TableHead>
+        <TableHead className="text-ink-muted text-right font-medium text-xs">סטטוס</TableHead>
+        <TableHead className="text-ink-muted text-right font-medium text-xs">נוצר</TableHead>
+        <TableHead className="text-navy" />
+      </TableRow>
+    </TableHeader>
+  );
+}
+
 function TableSkeleton({ cols }: { cols: number }) {
   return (
     <>
@@ -79,6 +97,38 @@ function TableSkeleton({ cols }: { cols: number }) {
   );
 }
 
+/** Accordion section meta — groups rows by discovery source (chair's request).
+ *  "צוטט ע״י דפנה" (corpus-decision reliance) is the high-value group and opens
+ *  by default; the noisy יומון group and the misc "אחר" group start collapsed. */
+type SectionKey = "chair" | "digest" | "other";
+const SECTION_META: Record<
+  SectionKey,
+  { label: string; title: string; desc: string; chip: string; defaultOpen: boolean }
+> = {
+  chair: {
+    label: "צוטט ע״י דפנה",
+    title: "פסיקה שהוועדה נסמכת עליה",
+    desc: "מצוטטת בהחלטות-הקורפוס (גרף-הציטוטים)",
+    chip: "bg-success-bg text-success",
+    defaultOpen: true,
+  },
+  digest: {
+    label: "יומון",
+    title: "זוהתה ביומון יומי",
+    desc: "מצביע, טרם צוטט בהחלטה",
+    chip: "bg-teal-bg text-teal",
+    defaultOpen: false,
+  },
+  other: {
+    label: "אחר",
+    title: "כתב-טענות / ידני",
+    desc: "צוטט בערר חי או נרשם ידנית",
+    chip: "bg-info-bg text-info",
+    defaultOpen: false,
+  },
+};
+const SECTION_ORDER: SectionKey[] = ["chair", "digest", "other"];
+
 type Props = {
   status?: MissingPrecedentStatus | "";
   q?: string;
@@ -91,7 +141,11 @@ export function MissingPrecedentsTable({ status, q, legalTopic }: Props) {
     status: status === "" ? undefined : status,
     q,
     legalTopic,
-    limit: 200,
+    // Load the full set so the accordion section counts (chair/digest/other)
+    // reflect the real totals — at limit:200 the page showed only the first
+    // page (e.g. 16 of 106 committee rows) while the header showed the true
+    // count, so the sections didn't sum to it. Backend caps at 2000.
+    limit: 1000,
   });
   const del = useDeleteMissingPrecedent();
 
@@ -108,6 +162,196 @@ export function MissingPrecedentsTable({ status, q, legalTopic }: Props) {
     }
   };
 
+  /* Partition the current result set by discovery source so each accordion
+   * section renders only its own rows. A row "cited by the committee" (has a
+   * resolved chair from the citation-graph bridge) wins over its raw
+   * discovery_source — that's the group the chair cares about. */
+  const groups = useMemo(() => {
+    const g: Record<SectionKey, MissingPrecedent[]> = { chair: [], digest: [], other: [] };
+    for (const mp of data?.items ?? []) {
+      if (mp.cited_by_chairs?.length) g.chair.push(mp);
+      else if (mp.discovery_source === "digest") g.digest.push(mp);
+      else g.other.push(mp);
+    }
+    return g;
+  }, [data]);
+
+  const renderRow = (mp: MissingPrecedent) => (
+    <TableRow
+      key={mp.id}
+      className="border-rule hover:bg-rule-soft/30 cursor-pointer"
+      onClick={() => setOpenId(mp.id)}
+    >
+      <TableCell className="max-w-[440px]">
+        {/* Single line when there's no distinct case_name — the old two-line
+            layout repeated the citation (name fell back to a truncation of the
+            same citation). Show the name row only when it adds information. */}
+        {mp.case_name && mp.case_name.trim() ? (
+          <>
+            <div className="text-sm text-navy font-semibold truncate">
+              {mp.case_name}
+            </div>
+            <div className="text-[0.72rem] text-ink-muted truncate" dir="rtl">
+              {mp.citation}
+            </div>
+          </>
+        ) : (
+          <div className="text-sm text-navy font-semibold truncate" dir="rtl">
+            {mp.citation}
+          </div>
+        )}
+      </TableCell>
+      <TableCell>
+        <span className="text-sm text-ink">{mp.legal_topic || "—"}</span>
+      </TableCell>
+      <TableCell>
+        {mp.cited_by_decisions?.length ? (
+          /* committee decision(s) that cite this missing ruling
+             (corpus citation-graph bridge) */
+          <span className="inline-flex items-baseline gap-1.5">
+            <span className="text-sm text-navy font-semibold tabular-nums" dir="ltr">
+              {mp.cited_by_decisions[0]}
+            </span>
+            {mp.cited_by_decisions.length > 1 ? (
+              <span className="text-[0.7rem] text-ink-muted">
+                +{mp.cited_by_decisions.length - 1}
+              </span>
+            ) : null}
+          </span>
+        ) : mp.cited_in_case_number ? (
+          <Link
+            href={`/cases/${encodeURIComponent(mp.cited_in_case_number)}`}
+            onClick={(e) => e.stopPropagation()}
+            className="text-sm text-navy hover:text-gold-deep inline-flex items-center gap-1"
+          >
+            {mp.cited_in_case_number}
+            <ExternalLink className="w-3 h-3" />
+          </Link>
+        ) : (
+          <span className="text-ink-muted text-sm">—</span>
+        )}
+      </TableCell>
+      <TableCell className="text-sm text-ink">
+        {mp.cited_by_chairs?.length ? (
+          /* cited by a committee DECISION in the corpus → show the chair who
+             relied on it (e.g. דפנה תמיר). Bridge from
+             precedent_internal_citations; takes priority over the generic
+             discovery-source chips. */
+          <>
+            <Badge
+              variant="outline"
+              className="rounded-full whitespace-nowrap bg-success-bg text-success border-transparent"
+            >
+              {mp.cited_by_chairs[0]}
+            </Badge>
+            {mp.cited_by_chairs.length > 1 ? (
+              <div className="text-[0.7rem] text-ink-muted mt-1">
+                +{mp.cited_by_chairs.length - 1} יו״ר
+              </div>
+            ) : null}
+          </>
+        ) : mp.discovery_source === "cited_only" ? (
+          <>
+            <Badge
+              variant="outline"
+              className="rounded-full whitespace-nowrap bg-plum-bg text-plum border-transparent"
+            >
+              פסיקה בקורפוס
+            </Badge>
+            {mp.cited_by_precedents?.length ? (
+              <div className="text-[0.7rem] text-ink-muted truncate max-w-[180px] mt-1">
+                מצוטט ע״י: {mp.cited_by_precedents.join(", ")}
+              </div>
+            ) : null}
+          </>
+        ) : mp.discovery_source === "digest" ? (
+          <>
+            <Badge
+              variant="outline"
+              className="rounded-full whitespace-nowrap bg-teal-bg text-teal border-transparent"
+            >
+              יומון
+            </Badge>
+            {mp.yomon_number ? (
+              <div className="text-[0.7rem] text-ink-muted mt-1">
+                מס׳ {mp.yomon_number}
+              </div>
+            ) : null}
+          </>
+        ) : (
+          <>
+            <SourceChip party={mp.cited_by_party} />
+            {mp.cited_by_party_name ? (
+              <div className="text-[0.7rem] text-ink-muted truncate max-w-[160px] mt-1">
+                {mp.cited_by_party_name}
+              </div>
+            ) : null}
+          </>
+        )}
+      </TableCell>
+      <TableCell>
+        <StatusBadge status={mp.status} />
+        {mp.linked_case_law_number ? (
+          <div className="text-[0.7rem] text-success mt-1">
+            ↳ {mp.linked_case_law_name || mp.linked_case_law_number}
+          </div>
+        ) : null}
+      </TableCell>
+      <TableCell className="text-[0.78rem] text-ink-muted">
+        {formatDate(mp.created_at)}
+      </TableCell>
+      <TableCell className="text-end">
+        <div className="flex items-center justify-end gap-2">
+          {mp.status === "open" ? (
+            /* gold "העלה והשלם" CTA (mockup 09 `.btn`) */
+            <Button
+              size="sm"
+              onClick={(e) => {
+                e.stopPropagation();
+                setOpenId(mp.id);
+              }}
+              className="h-7 bg-gold text-white hover:bg-gold-deep border-transparent text-[0.78rem] font-semibold"
+            >
+              <Upload className="w-3.5 h-3.5 me-1" />
+              העלה והשלם
+            </Button>
+          ) : (
+            /* passive "done" label for non-open rows */
+            <span className="inline-flex items-center gap-1 rounded-md bg-rule-soft text-ink-muted text-[0.78rem] font-medium px-2.5 py-1">
+              {mp.status === "closed" ? "קושר" : STATUS_LABELS[mp.status]} ✓
+            </span>
+          )}
+          {mp.status !== "open" ? (
+            <Button
+              variant="ghost"
+              size="sm"
+              onClick={(e) => {
+                e.stopPropagation();
+                setOpenId(mp.id);
+              }}
+              title="פרטים"
+            >
+              <Pencil className="w-4 h-4" />
+            </Button>
+          ) : null}
+          <Button
+            variant="ghost"
+            size="sm"
+            onClick={(e) => {
+              e.stopPropagation();
+              handleDelete(mp);
+            }}
+            disabled={del.isPending}
+            className="text-danger hover:text-danger"
+            title="מחיקה"
+          >
+            <Trash2 className="w-4 h-4" />
+          </Button>
+        </div>
+      </TableCell>
+    </TableRow>
+  );
+
   if (error) {
     return (
       <div className="rounded bg-danger-bg border border-danger/40 px-6 py-4 text-danger text-center text-sm">
@@ -116,168 +360,61 @@ export function MissingPrecedentsTable({ status, q, legalTopic }: Props) {
     );
   }
 
-  return (
+  if (isPending) {
-    <>
+    return (
       <div className="rounded-lg border border-rule bg-surface shadow-sm overflow-hidden">
         <Table>
-          <TableHeader className="bg-parchment">
+          <TableHeaderRow />
-            <TableRow className="border-rule hover:bg-transparent">
-              <TableHead className="text-ink-muted text-right font-medium text-xs">פסיקה</TableHead>
-              <TableHead className="text-ink-muted text-right font-medium text-xs">נושא</TableHead>
-              <TableHead className="text-ink-muted text-right font-medium text-xs">תיק</TableHead>
-              <TableHead className="text-ink-muted text-right font-medium text-xs">צוטט ע״י</TableHead>
-              <TableHead className="text-ink-muted text-right font-medium text-xs">סטטוס</TableHead>
-              <TableHead className="text-ink-muted text-right font-medium text-xs">נוצר</TableHead>
-              <TableHead className="text-navy" />
-            </TableRow>
-          </TableHeader>
           <TableBody>
-            {isPending ? (
+            <TableSkeleton cols={COLS} />
-              <TableSkeleton cols={7} />
-            ) : !data?.items.length ? (
-              <TableRow className="border-rule">
-                <TableCell colSpan={7} className="text-center text-ink-muted py-8">
-                  אין פסיקות חסרות בקריטריונים הנוכחיים.
-                </TableCell>
-              </TableRow>
-            ) : (
-              data.items.map((mp) => (
-                <TableRow
-                  key={mp.id}
-                  className="border-rule hover:bg-rule-soft/30 cursor-pointer"
-                  onClick={() => setOpenId(mp.id)}
-                >
-                  <TableCell className="max-w-[440px]">
-                    <div className="text-sm text-navy font-semibold truncate">
-                      {mp.case_name || mp.citation.split(" ").slice(0, 6).join(" ")}
-                    </div>
-                    <div className="text-[0.72rem] text-ink-muted truncate" dir="rtl">
-                      {mp.citation}
-                    </div>
-                  </TableCell>
-                  <TableCell>
-                    <span className="text-sm text-ink">{mp.legal_topic || "—"}</span>
-                  </TableCell>
-                  <TableCell>
-                    {mp.cited_in_case_number ? (
-                      <Link
-                        href={`/cases/${encodeURIComponent(mp.cited_in_case_number)}`}
-                        onClick={(e) => e.stopPropagation()}
-                        className="text-sm text-navy hover:text-gold-deep inline-flex items-center gap-1"
-                      >
-                        {mp.cited_in_case_number}
-                        <ExternalLink className="w-3 h-3" />
-                      </Link>
-                    ) : (
-                      <span className="text-ink-muted text-sm">—</span>
-                    )}
-                  </TableCell>
-                  <TableCell className="text-sm text-ink">
-                    {mp.discovery_source === "cited_only" ? (
-                      <>
-                        <Badge
-                          variant="outline"
-                          className="rounded-full whitespace-nowrap bg-plum-bg text-plum border-transparent"
-                        >
-                          פסיקה בקורפוס
-                        </Badge>
-                        {mp.cited_by_precedents?.length ? (
-                          <div className="text-[0.7rem] text-ink-muted truncate max-w-[180px] mt-1">
-                            מצוטט ע״י: {mp.cited_by_precedents.join(", ")}
-                          </div>
-                        ) : null}
-                      </>
-                    ) : mp.discovery_source === "digest" ? (
-                      <>
-                        <Badge
-                          variant="outline"
-                          className="rounded-full whitespace-nowrap bg-teal-bg text-teal border-transparent"
-                        >
-                          יומון
-                        </Badge>
-                        {mp.yomon_number ? (
-                          <div className="text-[0.7rem] text-ink-muted mt-1">
-                            מס׳ {mp.yomon_number}
-                          </div>
-                        ) : null}
-                      </>
-                    ) : (
-                      <>
-                        <SourceChip party={mp.cited_by_party} />
-                        {mp.cited_by_party_name ? (
-                          <div className="text-[0.7rem] text-ink-muted truncate max-w-[160px] mt-1">
-                            {mp.cited_by_party_name}
-                          </div>
-                        ) : null}
-                      </>
-                    )}
-                  </TableCell>
-                  <TableCell>
-                    <StatusBadge status={mp.status} />
-                    {mp.linked_case_law_number ? (
-                      <div className="text-[0.7rem] text-success mt-1">
-                        ↳ {mp.linked_case_law_name || mp.linked_case_law_number}
-                      </div>
-                    ) : null}
-                  </TableCell>
-                  <TableCell className="text-[0.78rem] text-ink-muted">
-                    {formatDate(mp.created_at)}
-                  </TableCell>
-                  <TableCell className="text-end">
-                    <div className="flex items-center justify-end gap-2">
-                      {mp.status === "open" ? (
-                        /* gold "העלה והשלם" CTA (mockup 09 `.btn`) */
-                        <Button
-                          size="sm"
-                          onClick={(e) => {
-                            e.stopPropagation();
-                            setOpenId(mp.id);
-                          }}
-                          className="h-7 bg-gold text-white hover:bg-gold-deep border-transparent text-[0.78rem] font-semibold"
-                        >
-                          <Upload className="w-3.5 h-3.5 me-1" />
-                          העלה והשלם
-                        </Button>
-                      ) : (
-                        /* passive "done" label for non-open rows */
-                        <span className="inline-flex items-center gap-1 rounded-md bg-rule-soft text-ink-muted text-[0.78rem] font-medium px-2.5 py-1">
-                          {mp.status === "closed" ? "קושר" : STATUS_LABELS[mp.status]} ✓
-                        </span>
-                      )}
-                      {mp.status !== "open" ? (
-                        <Button
-                          variant="ghost"
-                          size="sm"
-                          onClick={(e) => {
-                            e.stopPropagation();
-                            setOpenId(mp.id);
-                          }}
-                          title="פרטים"
-                        >
-                          <Pencil className="w-4 h-4" />
-                        </Button>
-                      ) : null}
-                      <Button
-                        variant="ghost"
-                        size="sm"
-                        onClick={(e) => {
-                          e.stopPropagation();
-                          handleDelete(mp);
-                        }}
-                        disabled={del.isPending}
-                        className="text-danger hover:text-danger"
-                        title="מחיקה"
-                      >
-                        <Trash2 className="w-4 h-4" />
-                      </Button>
-                    </div>
-                  </TableCell>
-                </TableRow>
-              ))
-            )}
           </TableBody>
         </Table>
       </div>
+    );
+  }
+
+  if (!data?.items.length) {
+    return (
+      <div className="rounded-lg border border-rule bg-surface shadow-sm px-6 py-10 text-center text-ink-muted text-sm">
+        אין פסיקות חסרות בקריטריונים הנוכחיים.
+      </div>
+    );
+  }
+
+  return (
+    <>
+      <div className="space-y-3.5">
+        {SECTION_ORDER.map((key) => {
+          const items = groups[key];
+          if (!items.length) return null;
+          const meta = SECTION_META[key];
+          return (
+            <details
+              key={key}
+              open={meta.defaultOpen}
+              className="group rounded-lg border border-rule bg-surface shadow-sm overflow-hidden"
+            >
+              <summary className="flex items-center gap-3 px-4 py-3.5 bg-parchment cursor-pointer select-none list-none [&::-webkit-details-marker]:hidden">
+                <ChevronDown className="w-4 h-4 text-ink-muted transition-transform -rotate-90 group-open:rotate-0 shrink-0" />
+                <span className={`rounded-full px-2.5 py-0.5 text-[0.72rem] font-semibold whitespace-nowrap ${meta.chip}`}>
+                  {meta.label}
+                </span>
+                <span className="font-bold text-navy text-sm">{meta.title}</span>
+                <span className="hidden sm:inline text-ink-muted text-[0.78rem]">
+                  — {meta.desc}
+                </span>
+                <span className="ms-auto tabular-nums text-[0.8rem] text-ink-soft bg-surface border border-rule rounded-full px-2.5 py-0.5 font-semibold">
+                  {items.length}
+                </span>
+              </summary>
+              <Table>
+                <TableHeaderRow />
+                <TableBody>{items.map(renderRow)}</TableBody>
+              </Table>
+            </details>
+          );
+        })}
+      </div>
 
       <MissingPrecedentDetailDrawer
         id={openId}

web-ui/src/components/precedents/link-related-dialog.tsx

@@ -18,6 +18,7 @@ import {
   useLinkRelatedCase,
   useUnlinkRelatedCase,
   RelatedCase,
+  IncomingCitation,
 } from "@/lib/api/precedent-library";
 
 const LEVEL_LABELS: Record<string, string> = {
@@ -133,19 +134,28 @@ function LinkDialog({ caseId, currentRelated, open, onOpenChange }: DialogProps)
 type SectionProps = {
   caseId: string;
   related: RelatedCase[];
+  /** Decisions that cite THIS ruling — auto-detected from the citation graph
+   *  (read-only). Optional so existing call-sites stay valid. */
+  incoming?: IncomingCitation[];
 };
 
 /* Rail-styled citations card (mockup 08 side rail). Renders linked related
  * decisions as a navy-headed card with arrow-prefixed rows; keeps the full
- * link/unlink logic. Used in the precedent-detail side rail. */
+ * link/unlink logic. Auto-detected incoming citations (decisions that cite this
-export function RelatedCasesSection({ caseId, related }: SectionProps) {
+ * ruling) render in the SAME row style, read-only (no unlink). Used in the
+ * precedent-detail side rail. */
+export function RelatedCasesSection({ caseId, related, incoming = [] }: SectionProps) {
   const [dialogOpen, setDialogOpen] = useState(false);
+  // De-dupe: a decision already linked manually shouldn't appear twice.
+  const manualIds = new Set(related.map((r) => r.id));
+  const autoCites = incoming.filter((c) => !manualIds.has(c.id));
+  const total = related.length + autoCites.length;
 
   return (
     <div className="rounded-lg border border-rule bg-surface shadow-sm px-4 py-3.5 space-y-2.5">
       <div className="flex items-center justify-between gap-2">
         <h3 className="text-navy text-[0.92rem] font-semibold m-0">
-          ציטוטים מקושרים{related.length > 0 ? ` (${related.length})` : ""}
+          ציטוטים מקושרים{total > 0 ? ` (${total})` : ""}
         </h3>
         <Button
           variant="outline"
@@ -157,10 +167,46 @@ export function RelatedCasesSection({ caseId, related }: SectionProps) {
         </Button>
       </div>
 
-      {related.length === 0 ? (
+      {total === 0 ? (
         <p className="text-ink-muted text-[0.82rem] m-0">אין החלטות קשורות עדיין</p>
       ) : (
         <ul className="list-none p-0 m-0">
+          {autoCites.map((c) => (
+            <li
+              key={c.id}
+              className="flex items-start gap-2 py-2 border-b border-rule-soft last:border-b-0"
+            >
+              <span className="text-gold font-bold leading-6 shrink-0" aria-hidden>←</span>
+              <a
+                href={`/precedents/${c.id}`}
+                className="min-w-0 flex-1 group hover:opacity-90 transition-opacity"
+              >
+                <div className="text-[0.82rem] text-ink-soft leading-5 group-hover:text-navy">
+                  {c.case_name || c.case_number}
+                </div>
+                <div className="flex items-center gap-2 mt-0.5 flex-wrap">
+                  {c.precedent_level && (
+                    <Badge
+                      variant="outline"
+                      className={`text-[0.6rem] ${LEVEL_COLORS[c.precedent_level] ?? ""}`}
+                    >
+                      {LEVEL_LABELS[c.precedent_level] ?? c.precedent_level}
+                    </Badge>
+                  )}
+                  {(c.chair_name || c.court) && (
+                    <span className="text-[0.68rem] text-ink-muted truncate">
+                      {c.chair_name || c.court}
+                    </span>
+                  )}
+                  {c.date && (
+                    <span className="text-[0.68rem] text-ink-muted tabular-nums" dir="ltr">
+                      {c.date.slice(0, 10)}
+                    </span>
+                  )}
+                </div>
+              </a>
+            </li>
+          ))}
           {related.map((r) => (
             <li
               key={r.id}

web-ui/src/components/training/curator-portrait-panel.tsx

@@ -45,6 +45,15 @@ export function CuratorPortraitPanel() {
       <StatsCard />
       <RecentFindings />
 
+      <div className="bg-info-bg border border-info/40 rounded-lg px-4 py-3 flex items-start gap-2.5 text-[0.78rem] text-ink-soft leading-relaxed">
+        <span>🔒</span>
+        <div>
+          <b className="text-navy">שער-היו״ר (INV-LRN1 / G10):</b> כל הערוצים <b>מציעים</b> בלבד.
+          ממצא או לקח משפיע על הכתיבה רק אחרי שדפנה אישרה אותו. האוצֵר עצמו read-only על התוכן —
+          מזהה דפוסים ומגיש, לעולם לא משנה קבצים או סגנון בעצמו.
+        </div>
+      </div>
+
       <Tabs defaultValue="curator-prompt" dir="rtl">
         <TabsList className="bg-rule-soft/60">
           <TabsTrigger value="curator-prompt">פרומפט ה-Curator</TabsTrigger>
@@ -67,38 +76,127 @@ export function CuratorPortraitPanel() {
 
 // ── stats card ─────────────────────────────────────────────────────
 
+function SectionLabel({ children }: { children: React.ReactNode }) {
+  return (
+    <div className="text-[0.72rem] uppercase tracking-wider text-gold-deep font-bold mb-3">
+      {children}
+    </div>
+  );
+}
+
+/** One of the three learning channels — engine, backing store, count, flow gate. */
+function ChannelCard({
+  letter, title, engine, target, store, value, unit, flows, flowLabel, meta,
+}: {
+  letter: string; title: string; engine: string; target: string; store: string;
+  value: number; unit: string; flows: boolean; flowLabel: string; meta: string;
+}) {
+  return (
+    <Card className="bg-surface border-rule">
+      <CardContent className="px-[18px] py-4 flex flex-col gap-2.5">
+        <div className="flex items-baseline gap-2">
+          <span className="w-[22px] h-[22px] rounded-md flex items-center justify-center font-bold text-xs text-parchment shrink-0 bg-navy">
+            {letter}
+          </span>
+          <div>
+            <h3 className="text-[0.9rem] text-navy font-semibold leading-tight">{title}</h3>
+            <div className="text-[0.72rem] text-ink-muted">{engine}</div>
+          </div>
+        </div>
+        <div className="text-[0.72rem] text-ink-soft bg-rule-soft rounded-md px-2.5 py-1.5">
+          ← {target} <code className="text-[0.68rem] text-gold-deep">{store}</code>
+        </div>
+        <div className="flex items-baseline gap-1.5">
+          <span className="text-[1.9rem] leading-none font-bold text-navy tabular-nums">{value}</span>
+          <span className="text-[0.78rem] text-ink-muted">{unit}</span>
+        </div>
+        <span
+          className={`inline-flex items-center gap-1.5 rounded-full text-[0.72rem] font-semibold px-2.5 py-0.5 w-fit border ${
+            flows
+              ? "bg-success-bg text-success border-success/40"
+              : "bg-gold-wash text-gold-deep border-gold/40"
+          }`}
+        >
+          <span className="text-[0.5rem]">●</span>{flowLabel}
+        </span>
+        <div className="text-[0.72rem] text-ink-muted leading-relaxed">{meta}</div>
+      </CardContent>
+    </Card>
+  );
+}
+
 function StatsCard() {
   const { data, isPending } = useCuratorStats();
 
   if (isPending) {
     return (
-      <div className="grid grid-cols-2 md:grid-cols-4 gap-3">
+      <div className="space-y-4">
-        {[...Array(4)].map((_, i) => <Skeleton key={i} className="h-20 w-full" />)}
+        <Skeleton className="h-16 w-full" />
+        <div className="grid grid-cols-1 md:grid-cols-3 gap-4">
+          {[...Array(3)].map((_, i) => <Skeleton key={i} className="h-44 w-full" />)}
+        </div>
       </div>
     );
   }
   if (!data) return null;
 
+  const { distillation: a, panel: b, curator: c } = data.channels;
+
   return (
-    <div className="grid grid-cols-2 md:grid-cols-4 gap-3">
+    <div className="space-y-5">
-      <Kpi label="ממצאי curator" value={data.total_findings} icon={<Sparkles className="w-4 h-4" />} />
+      <div className="bg-gold-wash border border-gold rounded-lg px-5 py-4 flex items-start gap-3">
-      <Kpi label="החלטות שנסקרו" value={`${data.decisions_with_findings}/${data.decisions_total}`} icon={<FileText className="w-4 h-4" />} />
+        <span className="text-gold-deep text-xl leading-tight">★</span>
-      <Kpi label="ממצאים מאושרים (זורמים לכותב)" value={data.findings_approved} icon={<CheckCircle2 className="w-4 h-4" />} />
+        <p className="text-ink-soft text-sm leading-relaxed">
-      <Kpi label="ממוצע ממצאים להחלטה"
+          הלמידה זורמת לכותב ב<b className="text-gold-deep font-semibold">שלושה ערוצים נפרדים</b>, כל אחד
-        value={
+          מופעל כשדפנה מסמנת החלטה כסופית. הכרטיס מראה כמה כל ערוץ תרם ומה כבר עובר בפועל לכותב —
-          data.decisions_with_findings > 0
+          רק ידע ש<b className="text-gold-deep font-semibold">אושר בשער-היו״ר</b> משפיע על הטיוטות (INV-LRN1).
-            ? (data.total_findings / data.decisions_with_findings).toFixed(1)
+        </p>
-            : "—"
+      </div>
-        }
+
-        icon={<Brain className="w-4 h-4" />}
+      <div>
-      />
+        <SectionLabel>שלושת ערוצי-ההזנה לכותב</SectionLabel>
+        <div className="grid grid-cols-1 md:grid-cols-3 gap-4">
+          <ChannelCard
+            letter="א" title="דיסטילציה" engine="Claude · השוואת טיוטה↔סופי"
+            target="כללי-מתודולוגיה" store="appeal_type_rules"
+            value={a.items_total} unit="פריטים זורמים"
+            flows={a.items_total > 0} flowLabel="זורם לכותב"
+            meta={`${a.discussion_rules} כללי-דיון + ${a.transition_phrases} ביטויי-מעבר · אושרו ב-${a.pairs_folded} זוגות דרך טאב ״למידה״.`}
+          />
+          <ChannelCard
+            letter="ב" title="פאנל דו-סוכני" engine="DeepSeek + Gemini · הצבעה 2/2"
+            target="לקחי-סגנון" store="decision_lessons"
+            value={b.total} unit="לקחים"
+            flows={b.approved > 0} flowLabel={`${b.approved} אושרו · ${b.proposed} ממתינים`}
+            meta={`על פני ${b.decisions} החלטות · אישור per-החלטה בטאב הקורפוס.`}
+          />
+          <ChannelCard
+            letter="ג" title="ממצאי-אוצֵר" engine="האוצֵר · קריאת הסופי + דפוסים"
+            target="ממצאי-סגנון" store="source=curator"
+            value={c.total} unit="ממצאים"
+            flows={c.approved > 0} flowLabel={`${c.approved} אושרו · ${c.proposed} ממתינים`}
+            meta="3–5 דפוסים מתויגים להחלטה (סגנון/מבנה/לקסיקון/טבלאי) — נתפסים מבנית, לא רק כהערה."
+          />
+        </div>
+      </div>
+
+      <div>
+        <SectionLabel>ערוץ האוצֵר — מבט מקרוב</SectionLabel>
+        <div className="grid grid-cols-2 md:grid-cols-4 gap-3">
+          <Kpi label="סך ממצאי-אוצֵר" value={c.total} icon={<Sparkles className="w-4 h-4" />} />
+          <Kpi label="החלטות שנסקרו" value={`${c.decisions_reviewed}/${data.decisions_total}`} icon={<FileText className="w-4 h-4" />} />
+          <Kpi label="מאושרים · זורמים לכותב" value={c.approved} tone="success" icon={<CheckCircle2 className="w-4 h-4" />} />
+          <Kpi label="ממתינים לאישור" value={c.proposed} tone="gold" icon={<Brain className="w-4 h-4" />} />
+        </div>
+      </div>
     </div>
   );
 }
 
 function Kpi({
-  label, value, icon,
+  label, value, icon, tone,
-}: { label: string; value: string | number; icon: React.ReactNode }) {
+}: { label: string; value: string | number; icon: React.ReactNode; tone?: "success" | "gold" }) {
+  const valueCls = tone === "success" ? "text-success" : tone === "gold" ? "text-gold-deep" : "text-navy";
   return (
     <Card className="bg-surface border-rule">
       <CardContent className="px-4 py-3">
@@ -106,7 +204,7 @@ function Kpi({
           {icon}
           <span>{label}</span>
         </div>
-        <p className="text-2xl text-navy font-semibold tabular-nums mt-1">{value}</p>
+        <p className={`text-2xl font-semibold tabular-nums mt-1 ${valueCls}`}>{value}</p>
       </CardContent>
     </Card>
   );
@@ -123,10 +221,10 @@ function RecentFindings() {
   if (!data || data.recent_findings.length === 0) {
     return (
       <Card className="bg-rule-soft/40 border-rule">
-        <CardContent className="px-6 py-5 text-center text-ink-muted text-sm">
+        <CardContent className="px-6 py-5 text-center text-ink-muted text-sm leading-relaxed">
-          אין עדיין ממצאים של ה-Curator. הוא מופעל אוטומטית כאשר דפנה מסמנת
+          אין עדיין ממצאי-אוצֵר. כשדפנה מסמנת החלטה כסופית, האוצֵר קורא את הסופי, מזהה
-          החלטה כסופית (mark-final), ושומר את ממצאיו כ-decision_lessons עם
+          3–5 דפוסי-סגנון, ושומר אותם כ-<code>decision_lessons</code> (source=&quot;curator&quot;)
-          source=&quot;curator&quot;.
+          הממתינים לאישורה — בנפרד מלקחי-הפאנל ומכללי-המתודולוגיה שלמעלה.
         </CardContent>
       </Card>
     );

web-ui/src/components/training/style-report-panel.tsx

@@ -197,7 +197,7 @@ export function StyleReportPanel() {
             <h3 className="text-navy text-lg m-0">אוצֵר — ממצאי-סגנון</h3>
             <div className="flex items-center gap-4 rounded-lg border border-success bg-success-bg px-[18px] py-3.5">
               <span className="text-success font-bold text-[1.75rem] leading-none tabular-nums">
-                {curator.data ? curator.data.findings_approved : "—"}
+                {curator.data ? curator.data.channels.curator.approved : "—"}
               </span>
               <div className="min-w-0">
                 <b className="text-navy text-sm font-semibold block">
@@ -214,8 +214,8 @@ export function StyleReportPanel() {
                   {curator.data
                     ? Math.max(
                         0,
-                        curator.data.total_findings -
+                        curator.data.channels.curator.total -
-                          curator.data.findings_approved,
+                          curator.data.channels.curator.approved,
                       )
                     : "—"}
                 </div>
@@ -223,7 +223,7 @@ export function StyleReportPanel() {
               </div>
               <div className="flex-1 rounded-lg border border-rule bg-rule-soft px-3.5 py-3">
                 <div className="text-ink-muted font-bold text-[1.35rem] tabular-nums leading-none">
-                  {curator.data ? curator.data.total_findings : "—"}
+                  {curator.data ? curator.data.channels.curator.total : "—"}
                 </div>
                 <div className="text-[0.78rem] text-ink-soft mt-1">סך ממצאים</div>
               </div>

web-ui/src/lib/api/agents.ts

@@ -179,3 +179,25 @@ export function useSubmitInteraction(caseNumber: string | undefined) {
     },
   });
 }
+
+export type AgentResetResult = {
+  ok: boolean;
+  reassigned_issues: { id: string; identifier: string }[];
+  reset_agents: { id: string; name: string; ok: boolean; error?: string }[];
+};
+
+export function useResetCaseAgents(caseNumber: string | undefined) {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: () =>
+      apiRequest<AgentResetResult>(
+        `/api/cases/${caseNumber}/agents/reset`,
+        { method: "POST" },
+      ),
+    onSuccess: () => {
+      if (caseNumber) {
+        qc.invalidateQueries({ queryKey: agentKeys.activity(caseNumber) });
+      }
+    },
+  });
+}

web-ui/src/lib/api/case-status.ts

@@ -74,6 +74,7 @@ const LEGACY_STATUS_LABELS: Record<string, string> = {
   ready_for_writing: "מוכן לכתיבה",
   drafting:          "בכתיבה",
   qa_failed:         "בדיקת איכות נכשלה",
+  qa_passed:         "טיוטה",
 };
 
 /**

web-ui/src/lib/api/cases.ts

@@ -176,14 +176,16 @@ export function useUpdateCase(caseNumber: string | undefined) {
         body: input,
       }),
     onSuccess: (data) => {
-      /* Patch cached detail and nudge the list to refetch on next focus */
+      /* Patch cached detail, then invalidate only the LIST queries — not
+       * casesKeys.all, which would also invalidate the detail we just patched
+       * and throw away the optimistic merge. */
       if (caseNumber) {
         qc.setQueryData<CaseDetail | undefined>(
           casesKeys.detail(caseNumber),
           (prev) => (prev ? { ...prev, ...data } : prev),
         );
       }
-      qc.invalidateQueries({ queryKey: casesKeys.all });
+      qc.invalidateQueries({ queryKey: [...casesKeys.all, "list"] });
     },
   });
 }

web-ui/src/lib/api/citation-verification.ts

@@ -0,0 +1,103 @@
+/**
+ * Citation-verification domain (X11 Phase 2 / #154) — the compose "אימות פסיקה" tab.
+ *
+ * Per legal argument: in-corpus supporting precedents with the cumulative authority
+ * signal (cited_by — followed/distinguished), their verify state + chair note, and
+ * per-issue radar (unlinked digests). The chair verifies before the writer cites
+ * (INV-AH). Backed by GET/POST /api/cases/{n}/citation-verification[/verify].
+ */
+
+import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query";
+import { apiRequest } from "./client";
+
+export type CitedBy = {
+  total: number;
+  positive: number;
+  negative: number;
+  unclassified: number;
+  by_treatment: Record<string, number>;
+};
+
+export type SupportingPrecedent = {
+  case_law_id: string;
+  case_number: string;
+  case_name: string;
+  quote: string;
+  score: number;
+  cited_by: CitedBy;
+  attached_id: string | null;
+  verified: boolean;
+  chair_note: string;
+};
+
+export type RadarLead = {
+  digest_id: string;
+  yomon_number?: string | number | null;
+  headline: string;
+  underlying_citation: string;
+  underlying_court: string;
+  score: number;
+  missing_precedent_id: string | null;
+  missing_precedent_status: string | null;
+  action: "new_lead" | "gap_open" | "fetched" | "available_link" | string;
+  matched_issues?: string[];
+};
+
+export type ArgumentBlock = {
+  argument_id: string;
+  title: string;
+  legal_topic: string;
+  priority: string;
+  party: string;
+  supporting: SupportingPrecedent[];
+  radar: RadarLead[];
+};
+
+export type CitationVerificationView = {
+  status: string;
+  case_number: string;
+  arguments: ArgumentBlock[];
+  summary: {
+    arguments_total: number;
+    arguments_with_support: number;
+    verified: number;
+    radar_leads: number;
+  };
+};
+
+export type VerifyCitationBody = {
+  argument_id: string;
+  case_law_id?: string;
+  quote?: string;
+  citation?: string;
+  chair_note?: string | null;
+  verified: boolean;
+  attached_id?: string;
+};
+
+export function useCitationVerification(caseNumber: string | undefined) {
+  return useQuery({
+    queryKey: ["citation-verification", caseNumber ?? ""],
+    queryFn: ({ signal }) =>
+      apiRequest<CitationVerificationView>(
+        `/api/cases/${caseNumber}/citation-verification`,
+        { signal },
+      ),
+    enabled: Boolean(caseNumber),
+    staleTime: 10_000,
+  });
+}
+
+export function useVerifyCitation(caseNumber: string | undefined) {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: (body: VerifyCitationBody) =>
+      apiRequest(`/api/cases/${caseNumber}/citation-verification/verify`, {
+        method: "POST",
+        body,
+      }),
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: ["citation-verification", caseNumber ?? ""] });
+    },
+  });
+}

web-ui/src/lib/api/citations.ts

@@ -9,7 +9,8 @@ import { apiRequest } from "./client";
 export type LinkedCitation = {
   citation: string;
   cited_id: string;
-  case_name: string;
+  /** May be null/empty — the panel guards with `c.case_name && …`. */
+  case_name: string | null;
   court: string;
   precedent_level: string;
 };

web-ui/src/lib/api/legal-arguments.ts

@@ -71,12 +71,30 @@ export function useLegalArguments(caseNumber: string | undefined) {
   });
 }
 
-export type AggregateArgumentsResult = {
+/**
-  status: "started" | string;
+ * The aggregation runs on the legal-analyst agent (host-side, where the
-  case_number: string;
+ * `claude` CLI lives) — NOT inline in the FastAPI container. The endpoint
-  force: boolean;
+ * either delegates via a Paperclip wakeup (`queued`) or short-circuits on a
-  message: string;
+ * cheap in-container DB pre-check (`no_claims` / `exists`). `skipped` means
-};
+ * no analyst route was available; the chair can run the MCP tool manually.
+ */
+export type AggregateArgumentsResult =
+  | {
+      status: "queued";
+      sub_issue_id: string;
+      analyst_id: string;
+      main_issue_id: string;
+    }
+  | {
+      status: "no_claims" | "exists";
+      total: number;
+      message: string;
+    }
+  | {
+      status: "skipped";
+      reason: "no_api_key" | "no_analyst" | "no_issue" | string;
+      company_id?: string;
+    };
 
 export function useAggregateArguments(caseNumber: string | undefined) {
   const qc = useQueryClient();

web-ui/src/lib/api/missing-precedents.ts

@@ -56,6 +56,10 @@ export type MissingPrecedent = {
   discovery_source: string | null; // manual | cited_only | digest | court_fetch
   cited_by_precedents: string[] | null; // corpus precedents citing a cited_only stub
   yomon_number: string | null; // digest gap's source yomon number
+  // Bridge to the corpus citation graph: committee decisions (and their chairs)
+  // that cite this still-missing ruling — fills "צוטט ע״י <יו״ר>" + "תיק".
+  cited_by_chairs: string[] | null;
+  cited_by_decisions: string[] | null;
 };
 
 export type MissingPrecedentListResponse = {

web-ui/src/lib/api/precedent-library.ts

@@ -143,10 +143,24 @@ export type RelatedCase = {
   relation_type: string;
 };
 
+/** A decision that cites THIS ruling — auto-detected from the citation graph
+ *  (precedent_internal_citations incoming edges). Read-only (no unlink). */
+export type IncomingCitation = {
+  id: string;
+  case_number: string;
+  case_name: string;
+  court: string;
+  precedent_level: string;
+  chair_name: string;
+  date: string | null;
+  confidence: number | null;
+};
+
 export type PrecedentDetail = Precedent & {
   full_text: string;
   halachot: Halacha[];
   related_cases: RelatedCase[];
+  incoming_citations: IncomingCitation[];
 };
 
 export type SearchHit =

web-ui/src/lib/api/training.ts

@@ -302,14 +302,39 @@ export type CuratorFinding = {
   created_at: string;
 };
 
+// One decision_lessons channel split by the INV-LRN1 review gate.
+export type LessonChannelStats = {
+  total: number;
+  approved: number;
+  proposed: number;
+  rejected: number;
+};
+
+// The three honest learning channels that feed the writer (INV-IA2/IA5).
+// Earlier this was a single flat count of source='curator' — which nothing
+// wrote, so it always read 0 while the panel/methodology channels carried all
+// the real learning. See get_curator_stats (web/app.py).
 export type CuratorStats = {
-  total_findings: number;
-  decisions_with_findings: number;
   decisions_total: number;
-  // LRN-5 (INV-IA5): the count of curator lessons with review_status='approved'
+  channels: {
-  // (the real INV-LRN1 writer gate) — replaces the old findings_applied, which
+    // A — Claude draft↔final distillation → methodology overrides; flows now.
-  // counted the informative-only applied_to_skill flag.
+    distillation: {
-  findings_approved: number;
+      discussion_rules: number;
+      transition_phrases: number;
+      items_total: number;
+      pairs_folded: number;
+      pairs_total: number;
+    };
+    // B — DeepSeek+Gemini panel → decision_lessons; flows when chair-approved.
+    panel: LessonChannelStats & {
+      decisions: number;
+      by_practice_area: Record<string, number>;
+    };
+    // C — curator agent's qualitative findings → decision_lessons (source='curator').
+    curator: LessonChannelStats & {
+      decisions_reviewed: number;
+    };
+  };
   recent_findings: CuratorFinding[];
 };
 

web/agent_platform_port.py

@@ -52,10 +52,12 @@ from web.paperclip_client import (
     post_comment as pc_post_comment,
     reject_interaction as pc_reject_interaction,
     reset_agent_session as pc_reset_agent_session,
+    reset_case_agents as pc_reset_case_agents,
     respond_to_interaction as pc_respond_to_interaction,
     restore_project as pc_restore_project,
     update_project_name as pc_update_project_name,
     wake_analyst_for_appraiser_facts as pc_wake_analyst_for_appraiser_facts,
+    wake_analyst_for_argument_aggregation as pc_wake_analyst_for_argument_aggregation,
     wake_ceo_agent as pc_wake_ceo,
     wake_ceo_for_feedback_fold as pc_wake_ceo_for_feedback_fold,
     wake_curator_for_final as pc_wake_curator_for_final,
@@ -99,6 +101,7 @@ __all__ = [
     "pc_wake_curator_for_final",
     "pc_wake_for_precedent_extraction",
     "pc_wake_analyst_for_appraiser_facts",
+    "pc_wake_analyst_for_argument_aggregation",
     # comments / interactions
     "pc_post_comment",
     "pc_get_issue_comments",
@@ -112,4 +115,5 @@ __all__ = [
     "pc_get_run_events",
     "pc_cancel_run",
     "pc_reset_agent_session",
+    "pc_reset_case_agents",
 ]

web/app.py

@@ -72,9 +72,11 @@ from web.agent_platform_port import (
     pc_reject_interaction,
     pc_request,
     pc_reset_agent_session,
+    pc_reset_case_agents,
     pc_respond_to_interaction,
     pc_restore_project,
     pc_wake_analyst_for_appraiser_facts,
+    pc_wake_analyst_for_argument_aggregation,
     rename_case_project,
     pc_wake_ceo,
     pc_wake_ceo_for_feedback_fold,
@@ -1316,31 +1318,77 @@ async def get_style_analyzer_prompt():
 
 @app.get("/api/training/curator/stats")
 async def get_curator_stats():
-    """Cheap aggregate stats over decision_lessons + style_corpus.
+    """Aggregate over the THREE learning channels that feed the writer, for the
+    /training "אוצֵר" tab. Each channel is surfaced honestly with its real
+    backing store and how much of it actually reaches the writer:
 
-    Used by the Curator-Portrait tab to show "10 curator findings across 24
+      A. distillation — Claude draft↔final diff → methodology overrides
-    decisions". We deliberately keep this server-side and aggregate so the
+         (appeal_type_rules['_global']). Folded items flow to the writer now.
-    UI can render a single card without fanning out N queries.
+      B. panel — DeepSeek+Gemini 2/2 vote → decision_lessons
+         (source='panel:deepseek+gemini'). Flow to the writer only once the
+         chair approves (review_status='approved', INV-LRN1/G10).
+      C. curator — the agent's Opus-read qualitative findings → decision_lessons
+         (source='curator'). Same chair gate as the panel.
+
+    Earlier this endpoint counted ONLY source='curator' and reported it as the
+    whole story, so it read 0 while the panel/methodology channels carried all
+    the real learning. The card was measuring a column nothing wrote (the agent
+    posted findings as Paperclip comments, never as decision_lessons — a
+    spec↔impl drift since closed). It now reports all three (INV-IA2/IA5).
     """
     pool = await db.get_pool()
     async with pool.acquire() as conn:
-        total_lessons = await conn.fetchval(
+        total_corpus = await conn.fetchval("SELECT count(*) FROM style_corpus") or 0
-            "SELECT count(*) FROM decision_lessons WHERE source = 'curator'"
+
+        # ── Channel A — distillation → methodology overrides ──────────────
+        meth_rows = await conn.fetch(
+            "SELECT rule_category, coalesce(jsonb_array_length(rule_value), 0) AS n "
+            "FROM appeal_type_rules "
+            "WHERE appeal_type = '_global' "
+            "  AND rule_category IN ('discussion_rules', 'transition_phrases')"
         )
-        decisions_with_findings = await conn.fetchval(
+        meth = {r["rule_category"]: r["n"] for r in meth_rows}
+        discussion_rules = meth.get("discussion_rules", 0)
+        transition_phrases = meth.get("transition_phrases", 0)
+        pairs_folded = await conn.fetchval(
+            "SELECT count(*) FROM draft_final_pairs WHERE status = 'lessons_folded'"
+        ) or 0
+        pairs_total = await conn.fetchval("SELECT count(*) FROM draft_final_pairs") or 0
+
+        # ── Channels B & C — decision_lessons by source × review_status ───
+        ls_rows = await conn.fetch(
+            "SELECT source, review_status, count(*) AS n FROM decision_lessons "
+            "GROUP BY source, review_status"
+        )
+
+        def _channel(src_predicate) -> dict:
+            buckets = {"approved": 0, "proposed": 0, "rejected": 0}
+            for r in ls_rows:
+                if src_predicate(r["source"] or ""):
+                    st = r["review_status"] or "proposed"
+                    buckets[st] = buckets.get(st, 0) + r["n"]
+            buckets["total"] = sum(buckets.values())
+            return buckets
+
+        panel = _channel(lambda s: s.startswith("panel:"))
+        curator = _channel(lambda s: s == "curator")
+
+        panel_decisions = await conn.fetchval(
+            "SELECT count(DISTINCT style_corpus_id) FROM decision_lessons "
+            "WHERE source LIKE 'panel:%'"
+        ) or 0
+        curator_decisions = await conn.fetchval(
             "SELECT count(DISTINCT style_corpus_id) FROM decision_lessons "
             "WHERE source = 'curator'"
+        ) or 0
+        pa_rows = await conn.fetch(
+            "SELECT sc.practice_area, count(*) AS n "
+            "FROM decision_lessons dl JOIN style_corpus sc ON sc.id = dl.style_corpus_id "
+            "WHERE dl.source LIKE 'panel:%' GROUP BY sc.practice_area"
         )
-        total_corpus = await conn.fetchval("SELECT count(*) FROM style_corpus")
+        panel_by_practice = {(r["practice_area"] or "—"): r["n"] for r in pa_rows}
-        # LRN-5 (INV-IA5): count the *real* consumer-mapped gate — review_status
+
-        # 'approved' is what flows to the writer (INV-LRN1, #126). The old count
+        # Recent curator findings (channel C) — newest first
-        # of applied_to_skill was a KPI over an informative-only flag (LRN-1)
-        # that writes nowhere, so it reported adoption that never happened.
-        approved = await conn.fetchval(
-            "SELECT count(*) FROM decision_lessons "
-            "WHERE source = 'curator' AND review_status = 'approved'"
-        )
-        # Last 10 curator findings — newest first
         recent_rows = await conn.fetch(
             """
             SELECT dl.id, dl.lesson_text, dl.category, dl.review_status,
@@ -1353,11 +1401,27 @@ async def get_curator_stats():
             LIMIT 10
             """
         )
+
     return {
-        "total_findings": total_lessons or 0,
+        "decisions_total": total_corpus,
-        "decisions_with_findings": decisions_with_findings or 0,
+        "channels": {
-        "decisions_total": total_corpus or 0,
+            "distillation": {
-        "findings_approved": approved or 0,
+                "discussion_rules": discussion_rules,
+                "transition_phrases": transition_phrases,
+                "items_total": discussion_rules + transition_phrases,
+                "pairs_folded": pairs_folded,
+                "pairs_total": pairs_total,
+            },
+            "panel": {
+                **panel,
+                "decisions": panel_decisions,
+                "by_practice_area": panel_by_practice,
+            },
+            "curator": {
+                **curator,
+                "decisions_reviewed": curator_decisions,
+            },
+        },
         "recent_findings": [
             {
                 "id": str(r["id"]),
@@ -2454,41 +2518,78 @@ async def api_get_claims(case_number: str):
 # the FastAPI container it short-circuits with status="llm_unavailable".
 
 @app.post("/api/cases/{case_number}/aggregate-arguments")
-async def api_aggregate_arguments(
+async def api_aggregate_arguments(case_number: str, force: bool = False):
-    case_number: str,
+    """Queue claim→argument aggregation by waking the legal-analyst agent.
-    background_tasks: BackgroundTasks,
-    force: bool = False,
-):
-    """Aggregate raw claims into distinct legal arguments via Claude.
 
-    Runs as a BackgroundTask because the LLM pass can take 30-90 seconds.
+    The aggregation itself calls `claude_session.query_json()`, which shells
+    out to the local `claude` CLI — present on the agent host, **absent in
+    this FastAPI container**. Running it inline as a BackgroundTask (the old
+    behaviour) silently produced nothing, and on `force` it destructively
+    deleted the existing arguments *before* the doomed LLM call. So we
+    delegate to the analyst exactly like `extract-appraiser-facts`: create a
+    child Paperclip issue, assign it to the company's analyst, and trigger a
+    wakeup. The analyst runs the MCP tool locally and posts results.
+
+    Cheap in-container pre-checks (DB only, no LLM) short-circuit before any
+    agent is spun up:
+      - `no_claims`  — there are no raw claims to aggregate yet.
+      - `exists`     — arguments already computed and `force` is False.
+
+    Response shape:
+        {"status": "queued", "sub_issue_id", "analyst_id", "main_issue_id"}
+      or {"status": "no_claims"|"exists", "total", "message"}
+      or {"status": "skipped", "reason": "no_api_key"|"no_analyst"|"no_issue"}
     """
     case = await db.get_case_by_number(case_number)
     if not case:
         raise HTTPException(404, f"תיק {case_number} לא נמצא")
 
-    async def _run() -> None:
+    case_id = UUID(case["id"])
-        try:
+    pool = await db.get_pool()
-            from legal_mcp.services import argument_aggregator
+    async with pool.acquire() as conn:
-            result = await argument_aggregator.aggregate_claims_to_arguments(
+        claim_count = await conn.fetchval(
-                UUID(case["id"]), force=force,
+            "SELECT COUNT(*) FROM claims WHERE case_id = $1", case_id,
-            )
+        )
-            logger.info(
+        existing_args = await conn.fetchval(
-                "aggregate_arguments[%s] finished: %s",
+            "SELECT COUNT(*) FROM legal_arguments WHERE case_id = $1", case_id,
-                case_number, result,
+        )
-            )
-        except Exception as e:  # noqa: BLE001
-            logger.exception(
-                "aggregate_arguments[%s] failed: %s", case_number, e,
-            )
 
-    background_tasks.add_task(_run)
+    if not claim_count:
-    return {
+        return {
-        "status": "started",
+            "status": "no_claims",
-        "case_number": case_number,
+            "total": 0,
-        "force": force,
+            "message": (
-        "message": "Aggregation started in background. Poll /legal-arguments for results.",
+                "אין טענות גולמיות בתיק. הרץ קודם חילוץ טענות (extract_claims) "
-    }
+                "ואז חשב טיעונים."
+            ),
+        }
+    if existing_args and not force:
+        return {
+            "status": "exists",
+            "total": existing_args,
+            "message": (
+                f"כבר קיימים {existing_args} טיעונים מאוגדים. השתמש בכפתור "
+                "החישוב-מחדש כדי לחשב מחדש (מוחק ובונה מחדש)."
+            ),
+        }
+
+    # Route to the analyst of the correct company by case-number prefix.
+    prefix = case_number[:1]
+    company_id = (
+        PAPERCLIP_COMPANIES["licensing"] if prefix == "1"
+        else PAPERCLIP_COMPANIES["betterment"] if prefix in ("8", "9")
+        else ""
+    )
+
+    try:
+        result = await pc_wake_analyst_for_argument_aggregation(
+            case_number, company_id=company_id, force=force,
+        )
+    except Exception as e:
+        logger.exception("analyst wakeup failed for argument aggregation %s", case_number)
+        raise HTTPException(500, f"לא ניתן לשלוח לאנליטיקאי: {e}")
+
+    return result
 
 
 @app.get("/api/cases/{case_number}/legal-arguments")
@@ -3155,6 +3256,62 @@ async def api_precedent_list(case_number: str):
     return envelope_unwrap(parsed)
 
 
+# ── Citation-verification panel (X11 Phase 2 / #154) — "אימות פסיקה" tab ──
+
+
+@app.get("/api/cases/{case_number}/citation-verification")
+async def api_citation_verification(case_number: str):
+    """Per legal-argument: supporting corpus precedents (with cited_by authority),
+    their verify state + chair note, and per-issue radar (unlinked digests). Powers
+    the compose 'אימות פסיקה' tab — the chair verifies before the writer cites."""
+    from legal_mcp.services import case_citation_verification as ccv
+    view = await ccv.build_view(case_number)
+    if view.get("status") == "case_not_found":
+        raise HTTPException(404, f"תיק {case_number} לא נמצא")
+    return view
+
+
+class CitationVerifyRequest(BaseModel):
+    argument_id: str
+    case_law_id: str = ""
+    quote: str = ""
+    citation: str = ""
+    chair_note: str | None = None
+    verified: bool = True
+    attached_id: str = ""   # set when the precedent row already exists
+
+
+@app.post("/api/cases/{case_number}/citation-verification/verify")
+async def api_citation_verify(case_number: str, req: CitationVerifyRequest):
+    """Verify / un-verify a precedent for a specific argument (the INV-AH gate the
+    writer respects). Upsert: PATCH an existing attachment, else attach a new one
+    linked to the argument + corpus ruling and mark it."""
+    case = await db.get_case_by_number(case_number)
+    if not case:
+        raise HTTPException(404, f"תיק {case_number} לא נמצא")
+    try:
+        if req.attached_id:
+            row = await db.set_case_precedent_verified(
+                UUID(req.attached_id), req.verified, req.chair_note)
+            if not row:
+                raise HTTPException(404, "שיוך-פסיקה לא נמצא")
+            return row
+        if not req.quote.strip() or not req.citation.strip():
+            raise HTTPException(400, "quote ו-citation חובה לצירוף חדש")
+        cid = case["id"]
+        row = await db.create_case_precedent(
+            case_id=UUID(cid) if isinstance(cid, str) else cid,
+            quote=req.quote, citation=req.citation,
+            chair_note=req.chair_note or "",
+            argument_id=UUID(req.argument_id) if req.argument_id else None,
+            case_law_id=UUID(req.case_law_id) if req.case_law_id else None,
+            verified=req.verified,
+        )
+        return row
+    except ValueError:
+        raise HTTPException(400, "מזהה לא תקין")
+
+
 @app.delete("/api/precedents/{precedent_id}")
 async def api_precedent_delete(precedent_id: str):
     """Delete a precedent attachment. The archived PDF (if any) stays
@@ -3548,9 +3705,11 @@ async def _enroll_final_in_library(
         out["error"] = "no final text extracted"
         return out
 
-    # Deterministic metadata from the case record — the Gemini metadata extractor is
+    # Deterministic seeds from the case record first — proceeding_type / date /
-    # tuned for EXTERNAL rulings and returns no_metadata for internal decisions, so we
+    # citation are identity, not inference, so we set them ourselves (no LLM) and the
-    # populate proceeding_type / date / tags / summary / citation ourselves (no LLM).
+    # later Gemini pass (reextract_metadata, below) is fills-empty-only and won't
+    # touch them. subject_tags/summary seeded here from the case (chair-curated
+    # subject_categories win when present); when empty, Gemini fills them inline.
     district = "ירושלים"
     proceeding_type = (case.get("proceeding_type") or "ערר").strip()
     decision_date = case.get("decision_date") or case.get("hearing_date")
@@ -3590,6 +3749,42 @@ async def _enroll_final_in_library(
     except Exception as e:
         logger.warning("citation build failed for %s: %s", case_number, e)
 
+    # Fill the LLM-derived metadata (subject_tags, summary, headnote, key_quote)
+    # inline via Gemini (REST — container-safe; GOOGLE_GEMINI_API_KEY in Coolify),
+    # reusing the same reextract_metadata path the UI/drain use (G2 — one path, full
+    # status lifecycle). apply_to_record fills ONLY empty fields, so the deterministic
+    # seeds above (proceeding_type / date / citation, plus any chair-curated
+    # subject_categories) are preserved. Without this the row sat at
+    # metadata_extraction_status='pending' with empty tags until a drain happened to
+    # pick it up, so a freshly-enrolled final showed no subject tags in the edit UI.
+    # Best-effort — surfaced in the return value, never silently swallowed.
+    try:
+        from legal_mcp.services import precedent_library as plib_service
+        meta = await plib_service.reextract_metadata(UUID(case_law_id))
+        out["metadata"] = {
+            "status": meta.get("status"),
+            "fields": meta.get("fields") or [],
+        }
+    except Exception as e:
+        logger.warning("inline metadata extraction failed for %s: %s", case_number, e)
+        out["metadata_error"] = str(e)
+
+    # Grow the style-exemplar corpus (channel B): break this final into block-level
+    # paragraphs the writer retrieves (07-learning §0.2). Before this, exemplars were
+    # frozen at the one-time seed backfill — new finals never got exemplarized, so the
+    # richest style channel never grew. Same extraction the backfill uses (G2). Voyage
+    # embeds over REST → container-safe. Best-effort; surfaced, never fails the upload.
+    try:
+        from legal_mcp.services import style_exemplars as _sx
+        n_ex = await _sx.extract_and_store(
+            decision_number=case_number, source="internal_committee",
+            full_text=final_text, practice_area=case.get("practice_area", ""),
+        )
+        out["exemplars"] = n_ex
+    except Exception as e:
+        logger.warning("style-exemplar extraction failed for %s: %s", case_number, e)
+        out["exemplars_error"] = str(e)
+
     # The precedents this decision cites → link to the library; flag the ones not found.
     try:
         await cit_tools.extract_internal_citations(case_law_id=case_law_id, limit=0)
@@ -3718,6 +3913,25 @@ async def api_upload_final_decision(case_number: str, file: UploadFile = File(..
     # checking + missing-precedent flagging (07-learning §1.3). Surfaced, not silent.
     library = await _enroll_final_in_library(case, case_number, final_text, chair_name)
 
+    # Path A — prospective held-out snapshot. Right now the draft (decision_blocks)
+    # was written with only the PRIOR lesson pool — this case's lessons fold later,
+    # manually, in /training. So measuring style-distance HERE is a clean
+    # generalization datapoint. Append-only; best-effort (never fails the upload).
+    try:
+        from legal_mcp.services import style_distance as _sd
+        sd = await _sd.style_distance(case_number)
+        if isinstance(sd, dict) and "error" not in sd:
+            pool = await db.voice_lesson_pool_sizes()
+            summ = sd.get("summary", {})
+            await db.record_style_distance_snapshot(
+                case_number, pair_id,
+                pool.get("discussion_rules", 0), pool.get("transition_phrases", 0),
+                summ.get("anti_pattern_total"), summ.get("ratio_max_deviation_pp"),
+                summ.get("change_percent"),
+            )
+    except Exception as e:
+        logger.warning("held-out style-distance snapshot failed for %s: %s", case_number, e)
+
     case_dir = config.find_case_dir(case_number)
     if case_dir.exists():
         commit_and_push(case_dir, f"החלטה סופית של היו\"ר: {final_name}")
@@ -4079,6 +4293,19 @@ async def api_post_interaction_response(
         raise HTTPException(502, f"שגיאת Paperclip: {e}")
 
 
+@app.post("/api/cases/{case_number}/agents/reset")
+async def api_reset_case_agents(case_number: str):
+    """Reset stuck agents for a case.
+
+    Clears writer/QA agents from 'error' status and reassigns any open
+    issues back to the chair user, stopping Paperclip recovery loops.
+    """
+    result = await pc_reset_case_agents(case_number)
+    if not result.get("ok"):
+        raise HTTPException(502, result.get("error", "שגיאה בביצוע האיפוס"))
+    return result
+
+
 # ── Settings: MCP Server Configuration ────────────────────────────
 #
 # Source of truth for legal-ai env vars is Coolify (see memory:
@@ -4630,6 +4857,15 @@ async def api_learning_style_distance(case_number: str):
     return await _sd.style_distance(case_number)
 
 
+@app.get("/api/learning/style-distance-history")
+async def api_learning_style_distance_history():
+    """Path A — מגמת ה-held-out הפרוספקטיבי: snapshot של מרחק-הסגנון שנלכד בכל
+    העלאת-סופי *לפני* הטמעת-לקחי-התיק, מול גודל-בריכת-הלקחים באותו רגע. ירידה
+    ב-anti_pattern_total / change_percent ככל שהבריכה גדלה = הלמידה מכלילה."""
+    items = await db.get_style_distance_history()
+    return {"items": items, "count": len(items)}
+
+
 def _coerce_json(raw):
     if isinstance(raw, str):
         try:
@@ -4670,36 +4906,13 @@ class PromoteLearningRequest(BaseModel):
 
 
 async def _append_methodology_override(category: str, key: str, items: list[str]) -> None:
-    """Read current (override-or-default) list value, append new items, upsert override.
+    """Thin wrapper over db.append_global_rule (the single locked append impl, G2) —
-    Shared by the T14 approval gate to fold approved learnings into writer-consumed channels.
+    seeds from _METHODOLOGY_DEFAULTS when no override row exists yet. Shared by the
-
+    T14 promote gate; chair-feedback auto-flow calls db.append_global_rule directly."""
-    MET-2/3 (INV-IA3): the read-modify-write runs inside ONE transaction with the
+    await db.append_global_rule(
-    existing override row locked FOR UPDATE, so a concurrent promote (or a methodology
+        category, key, items,
-    PUT) can't interleave between the read and the write and silently drop items.
+        seed_if_missing=list(_METHODOLOGY_DEFAULTS.get(category, {}).get(key, [])),
-    The methodology PUT overwrites the same row; the MET-1 invalidation (גל-1) makes
+    )
-    the /methodology editor refetch on promote so it edits post-append state."""
-    pool = await db.get_pool()
-    async with pool.acquire() as conn:
-        async with conn.transaction():
-            row = await conn.fetchrow(
-                "SELECT rule_value FROM appeal_type_rules "
-                "WHERE appeal_type = '_global' AND rule_category = $1 AND rule_key = $2 "
-                "FOR UPDATE",
-                category, key,
-            )
-            if row:
-                current = _coerce_json(row["rule_value"]) or []
-            else:
-                current = list(_METHODOLOGY_DEFAULTS.get(category, {}).get(key, []))
-            if not isinstance(current, list):
-                current = []
-            merged = current + [s for s in items if s and s not in current]
-            await conn.execute(
-                "INSERT INTO appeal_type_rules (id, appeal_type, rule_category, rule_key, rule_value) "
-                "VALUES (gen_random_uuid(), '_global', $1, $2, $3::text::jsonb) "
-                "ON CONFLICT (appeal_type, rule_category, rule_key) DO UPDATE SET rule_value = $3::text::jsonb",
-                category, key, json.dumps(merged, ensure_ascii=False),
-            )
 
 
 @app.post("/api/learning/pairs/{pair_id}/promote")
@@ -6444,6 +6657,17 @@ class DigestLinkRequest(BaseModel):
     case_law_id: str
 
 
+@app.get("/api/cases/{case_number}/digest-radar")
+async def api_case_digest_radar(case_number: str, limit: int = 5, min_score: float = 0.45):
+    """Case-contextual digest radar (X12) — UNLINKED digests whose topic is close to
+    this case (rulings we don't hold yet). Powers the case-page "📡 רדאר יומונים" lead
+    so a relevant ruling known only via a digest doesn't fall through the cracks while
+    the case is decided. INV-DIG1: points at the underlying ruling, never cites the
+    digest."""
+    return await digest_service.case_digest_radar(
+        case_number, limit=max(1, min(int(limit), 20)), min_score=float(min_score))
+
+
 @app.post("/api/digests/upload")
 async def digest_upload(
     file: UploadFile = File(...),
@@ -7141,7 +7365,7 @@ async def internal_decisions_upload(
     practice_area: str = Form(""),
     appeal_subtype: str = Form(""),
     subject_tags: str = Form("[]"),
-    is_binding: bool = Form(True),
+    is_binding: bool = Form(False),  # INV-DM7: committee = persuasive (coerced at db too)
     summary: str = Form(""),
 ):
     """Upload a planning appeals-committee decision to the internal corpus.
@@ -7877,7 +8101,7 @@ async def missing_precedents_list(
         case_id=case_uuid,
         legal_topic=legal_topic.strip() or None,
         q=q.strip() or None,
-        limit=max(1, min(int(limit), 500)),
+        limit=max(1, min(int(limit), 2000)),
         offset=max(0, int(offset)),
     )
     # Counters useful for the sidebar badge.

web/paperclip_client.py

@@ -791,6 +791,71 @@ async def reset_agent_session(agent_id: str) -> dict:
     return resp.json()
 
 
+CHAIM_USER_ID = "ZpDWXxFweC3MftuF1Ttyu2VUbSPHbKZd"
+
+
+async def reset_case_agents(case_number: str) -> dict:
+    """Reset agent state for a case: clear error status + reassign stuck issues to user.
+
+    Two actions:
+    1. Any non-completed issue still assigned to an agent is reassigned to the chair user,
+       stopping Paperclip's source_scoped_recovery_action loop.
+    2. Every agent in the case's company whose global status is 'error' gets a
+       reset_agent_session call (clears wedged runtime) and its DB status set to 'idle'.
+    """
+    first_digit = case_number.split("-")[0][0] if case_number else "1"
+    company_id = COMPANIES["betterment"] if first_digit in ("8", "9") else COMPANIES["licensing"]
+
+    conn = await asyncpg.connect(PAPERCLIP_DB_URL)
+    try:
+        project = await conn.fetchrow(
+            "SELECT id FROM projects WHERE name LIKE $1 LIMIT 1",
+            f"%{case_number}%",
+        )
+        if not project:
+            return {"ok": False, "error": f"No Paperclip project found for {case_number}"}
+
+        project_id = project["id"]
+
+        reassigned = await conn.fetch(
+            """UPDATE issues
+               SET assignee_agent_id = null, assignee_user_id = $1, updated_at = now()
+               WHERE project_id = $2
+                 AND assignee_agent_id IS NOT NULL
+                 AND status NOT IN ('done', 'cancelled')
+               RETURNING id, identifier""",
+            CHAIM_USER_ID, project_id,
+        )
+
+        error_agents = await conn.fetch(
+            "SELECT id, name FROM agents WHERE company_id = $1::uuid AND status = 'error'",
+            company_id,
+        )
+
+        if error_agents:
+            await conn.execute(
+                "UPDATE agents SET status = 'idle' WHERE id = ANY($1::uuid[])",
+                [r["id"] for r in error_agents],
+            )
+    finally:
+        await conn.close()
+
+    reset_results = []
+    for agent in error_agents:
+        try:
+            await reset_agent_session(str(agent["id"]))
+            reset_results.append({"id": str(agent["id"]), "name": agent["name"], "ok": True})
+        except Exception as e:
+            logger.warning("reset_agent_session failed for %s: %s", agent["id"], e)
+            reset_results.append({"id": str(agent["id"]), "name": agent["name"], "ok": False, "error": str(e)})
+
+    return {
+        "ok": True,
+        "reassigned_issues": [{"id": str(r["id"]), "identifier": r["identifier"]} for r in reassigned],
+        "reset_agents": reset_results,
+    }
+
+
 async def respond_to_interaction(
     issue_id: str, interaction_id: str, payload: dict,
 ) -> dict:
@@ -1371,3 +1436,121 @@ async def wake_analyst_for_appraiser_facts(
         "analyst_id": analyst_id,
         "main_issue_id": main_issue_id,
     }
+
+
+async def wake_analyst_for_argument_aggregation(
+    case_number: str,
+    company_id: str,
+    force: bool = False,
+) -> dict:
+    """Wake the legal-analyst to aggregate raw claims into legal arguments.
+
+    Triggered by the chair clicking "חשב טיעונים" in the case page. The
+    FastAPI container cannot run `aggregate_claims_to_arguments` directly —
+    the aggregator calls `claude_session.query_json()`, which only works
+    where the local `claude` CLI is present (the MCP server / agent runner
+    on the host), **not in this container**. So instead of an in-container
+    BackgroundTask (which fails silently and, on `force`, destructively
+    deletes the existing arguments before the doomed LLM call), we create a
+    child issue under the case's main Paperclip issue, assign it to the
+    analyst of the correct company, and trigger a wakeup. The analyst's
+    HEARTBEAT picks up the issue, runs the MCP tool locally, and reports
+    back via a comment.
+
+    Mirrors ``wake_analyst_for_appraiser_facts`` — same delegation shape.
+
+    Returns a dict shaped for the FastAPI endpoint to serialize as-is:
+        {"status": "queued", "sub_issue_id", "analyst_id", "main_issue_id"}
+      or {"status": "skipped", "reason": "..."} for non-fatal early outs.
+    """
+    if not PAPERCLIP_BOARD_API_KEY:
+        logger.warning(
+            "PAPERCLIP_BOARD_API_KEY not set — cannot queue analyst wakeup "
+            "for argument aggregation on %s",
+            case_number,
+        )
+        return {"status": "skipped", "reason": "no_api_key"}
+
+    analyst_id = ANALYST_AGENTS.get(company_id)
+    if not analyst_id:
+        logger.info("No analyst configured for company %s — skipping", company_id)
+        return {"status": "skipped", "reason": "no_analyst", "company_id": company_id}
+
+    issues = await get_case_issues(case_number)
+    if not issues:
+        logger.warning(
+            "No Paperclip issues found for case %s — cannot queue analyst", case_number,
+        )
+        return {"status": "skipped", "reason": "no_issue"}
+
+    main_issue = next((i for i in issues if i.get("status") == "in_progress"), None) or issues[0]
+    main_issue_id = main_issue["id"]
+
+    force_clause = ", force=True" if force else ""
+    rerun_note = (
+        "זהו חישוב-מחדש (force) — הטיעונים הקיימים יימחקו ויחושבו מחדש.\n\n"
+        if force else ""
+    )
+    description = (
+        f"חיים ביקש חישוב טיעונים משפטיים בתיק {case_number}.\n\n"
+        f"{rerun_note}"
+        f"הרץ `mcp__legal-ai__aggregate_claims_to_arguments(case_number=\"{case_number}\"{force_clause})` "
+        f"וכתוב comment בעברית עם תוצאת האיגוד — כמה טיעונים מובחנים נוצרו לכל צד "
+        f"(עוררים / משיבים / ועדה / מבקשי-היתר). אם אין טענות גולמיות בתיק, דווח "
+        f"ב-comment שצריך להריץ קודם חילוץ טענות (`extract_claims`) וסגור את ה-issue כ-blocked."
+    )
+    child_resp = await pc_request(
+        "POST",
+        f"/api/issues/{main_issue_id}/children",
+        json={
+            "title": f"[ערר {case_number}] חישוב טיעונים משפטיים",
+            "description": description,
+            "status": "in_progress",
+            # Paperclip ISSUE_PRIORITIES = critical|high|medium|low — "normal"
+            # is NOT a valid enum value (Zod 400 → surfaces as 500).
+            "priority": "medium",
+            "assigneeAgentId": analyst_id,
+        },
+        raise_on_error=True,
+    )
+    sub_issue = child_resp.json()
+    sub_issue_id = sub_issue["id"]
+
+    # Tag plugin_state so the case page surfaces this sub-issue too.
+    try:
+        conn = await asyncpg.connect(PAPERCLIP_DB_URL)
+        try:
+            await _link_case_to_issue(conn, sub_issue_id, case_number)
+        finally:
+            await conn.close()
+    except Exception as e:
+        logger.warning("plugin_state link failed for sub_issue=%s: %s", sub_issue_id, e)
+
+    wake_resp = await pc_request(
+        "POST",
+        f"/api/agents/{analyst_id}/wakeup",
+        json={
+            "source": "on_demand",
+            "triggerDetail": "manual",
+            "reason": f"aggregate_arguments_{case_number}",
+            # "assignment" is the generic mutation the HEARTBEAT recognises;
+            # task intent lives in the child-issue description, not the payload.
+            "payload": {
+                "issueId": sub_issue_id,
+                "mutation": "assignment",
+                "caseNumber": case_number,
+            },
+        },
+        raise_on_error=True,
+    )
+    logger.info(
+        "Analyst wakeup for argument aggregation on case %s: sub_issue=%s "
+        "analyst=%s force=%s wake=%s",
+        case_number, sub_issue_id, analyst_id, force, wake_resp.status_code,
+    )
+    return {
+        "status": "queued",
+        "sub_issue_id": sub_issue_id,
+        "analyst_id": analyst_id,
+        "main_issue_id": main_issue_id,
+    }