fix(chunker): תיקון זיהוי כותרות טענות הצדדים ביחיד/נקבה + שלוש שכבות הגנה

**סיבת-שורש:** רג'קס respondent_claims כיסה רק צורת רבים (המשיבים/המשיבין),
ולא יחיד נקבה (המשיבה) ויחיד זכר (המשיב). הכותרת "טענות המשיבה:" בתיק 8181-21
נבלעה לתוך מקטע ruling → חולצה כהלכה שגויה.

**שלוש שכבות הגנה:**
1. chunker.py — הרחבת SECTION_PATTERNS לכסות יחיד/זכר/נקבה + תגובת/תשובת
2. halacha_extractor.py — עיגון חיובי: drop ל-ruling chunks שלפני legal_analysis
3. halacha_quality.py — FLAG_PARTY_CLAIM: זיהוי שפת-טענות-צד בציטוט התומך

**היקף:** 93 תיקים עם 0 chunks של טענות (כנראה בגלל כותרות ביחיד שלא הוכרו);
628 הלכות מאושרות מתיקים אלה — חלקן עשויות להיות תקינות, יש לעשות re-chunk.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

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

@@ -279,9 +279,50 @@ FLAG_NON_DECISION = "non_decision"
 FLAG_TRUNCATED_QUOTE = "truncated_quote"
 FLAG_THIN_RESTATEMENT = "thin_restatement"
 FLAG_QUOTE_UNVERIFIED = "quote_unverified"
-FLAG_NLI_UNSUPPORTED = "nli_unsupported"  # rule not entailed by its quote (#81.3)
-FLAG_APPLICATION = "application"           # fact-dependent, not a holding (#81.4)
-FLAG_NEAR_DUPLICATE = "near_duplicate"     # cosine-tail lexical dup (#82.3)
+FLAG_NLI_UNSUPPORTED = "nli_unsupported"   # rule not entailed by its quote (#81.3)
+FLAG_APPLICATION = "application"            # fact-dependent, not a holding (#81.4)
+FLAG_NEAR_DUPLICATE = "near_duplicate"      # cosine-tail lexical dup (#82.3)
+FLAG_PARTY_CLAIM = "party_claim_language"   # quote reads as a party's position, not the court's
+
+
+# ── Party-claim language: quote is the court's words, not a party's ──
+#
+# Positive markers that a quote comes from a party's argument section rather
+# than the court's own reasoning.  The chunker now correctly classifies these
+# sections, but a belt-and-suspenders lexical gate catches any case where
+# the chunker still absorbs a party-claims section into a reasoning chunk
+# (e.g. an unrecognised header variant).  We scan the supporting_quote only —
+# the rule_statement is already abstracted and should not contain these phrases.
+
+_PARTY_CLAIM_MARKERS = (
+    # Named-party attribution forms — always party-claim language, never court reasoning
+    "לטענת העורר",
+    "לטענת העוררת",
+    "לטענת העוררים",
+    "לטענת המשיב",
+    "לטענת המשיבה",
+    "לטענת המשיבים",
+    "טוען העורר",
+    "טוענת העוררת",
+    "טוען המשיב",
+    "טוענת המשיבה",
+    # Excluded (too broad — courts also use these in their own reasoning):
+    # "נטען כי", "נטען על ידי", "נטען על-ידי", "לטענתו", "לטענתה", "לטענתם"
+)
+
+
+def detect_party_claim_language(supporting_quote: str) -> str | None:
+    """Return the first party-claim marker found in the quote (or None).
+
+    Only the supporting_quote is scanned — rule_statement is already abstracted.
+    A match means the LLM likely extracted from a party's argument section
+    rather than the court's reasoning.
+    """
+    norm = normalize_text(supporting_quote)
+    for marker in _PARTY_CLAIM_MARKERS:
+        if marker in norm:
+            return marker
+    return None
 
 
 # ── NLI entailment check (rule_statement ⊨ supporting_quote) — #81.3 ──
@@ -417,4 +458,8 @@ def compute_quality_flags(
     # rule_type='application' and add a high-precision deixis catch.
     if rule_type == "application" or is_fact_dependent(rule_statement):
         flags.append(FLAG_APPLICATION)
+    # Belt-and-suspenders: if the quote contains party-claim language the
+    # chunker's section filter should have excluded, flag for manual review.
+    if detect_party_claim_language(supporting_quote):
+        flags.append(FLAG_PARTY_CLAIM)
     return flags