feat(halacha): #82.4 provenance-union on dedup-skip + #82.6 over-merge guard

חילוץ החלטת-ה-dedup ל-helper טהור ובדיק `halacha_quality.dedup_action()` (skip/flag/keep),
ושני שיפורים על מסלול ה-dedup-on-insert:

#82.4 — merge-with-provenance, לא blind-drop: כשמדלגים על כפילות-סמנטית (cosine≥0.93),
מאחדים את ה-`cites` של השורה הנכנסת אל השכן הקנוני ששורד (במקום לאבד אותם). זהו שדה-ה-
provenance היחיד שקיים בהכנסה; בחירת-קנוני + מיזוג-corroboration מלא שייכים למסלול ה-
reconimation הלא-מקוון (#82.7 / #84.2, שם לשורות כבר יש provenance מצטבר) — מתועד בקוד.

#82.6 — over-merge guard: ההחלטה PAIRWISE מול שכן יחיד הקרוב ביותר, ורק השורה הנכנסת
מודלגת אי-פעם (אף שורה קיימת לא ממוזגת/נמחקת). אין connected-components closure בהכנסה,
לכן שרשרת A~B~C לא קורסת לשורה אחת גם כש-A,C מובחנים. מתועד ב-dedup_action + נבדק.

invariants: G1 (provenance נשמר במקור, לא אובד) · G2 (לוגיקת-החלטה ב-helper יחיד בדיק,
refactor משמר-התנהגות) · INV-G10 (אין auto-merge של שורות קיימות; tail→flag→סקירת-יו"ר).
tests: 6 חדשות (skip/flag/keep/over-merge/boundaries) + 59 בדיקות-הלכה קיימות עוברות.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

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

@@ -4299,22 +4299,39 @@ async def store_halachot_for_chunk(
                 flags = list(h.get("quality_flags") or [])
                 if emb is not None and config.HALACHA_DEDUP_COSINE <= 1.0:
                     neighbor = await conn.fetchrow(
-                        "SELECT rule_statement, (embedding <=> $2) AS dist "
+                        "SELECT id, rule_statement, cites, (embedding <=> $2) AS dist "
                         "FROM halachot WHERE case_law_id = $1 "
                         "AND embedding IS NOT NULL "
                         "ORDER BY embedding <=> $2 LIMIT 1",
                         case_law_id, emb,
                     )
                     if neighbor is not None:
-                        dist = float(neighbor["dist"])
-                        if dist <= dedup_distance:
+                        # PAIRWISE decision vs the single nearest neighbor — no
+                        # cluster closure, so a chain A~B~C can't over-merge to one
+                        # row (#82.6 over-merge guard). See halacha_quality.dedup_action.
+                        action = halacha_quality.dedup_action(
+                            float(neighbor["dist"]), h["rule_statement"],
+                            neighbor["rule_statement"], dedup_distance, band_distance,
+                        )
+                        if action == "skip":
+                            # #82.4 — merge-with-provenance, not blind drop: fold the
+                            # incoming row's cites into the surviving neighbor (the
+                            # only provenance present at insert; full canonical-
+                            # selection/merge lives in the offline reconciliation
+                            # path, #82.7 / #84.2).
+                            new_cites = [c for c in (h.get("cites") or []) if c]
+                            if new_cites:
+                                await conn.execute(
+                                    "UPDATE halachot SET cites = ARRAY(SELECT DISTINCT "
+                                    "unnest(COALESCE(cites, '{}') || $2::text[])), "
+                                    "updated_at = now() WHERE id = $1",
+                                    neighbor["id"], new_cites,
+                                )
                             skipped += 1
                             continue
                         # tail band: below auto-skip but lexically near → flag.
-                        if (dist <= band_distance
-                                and halacha_quality.FLAG_NEAR_DUPLICATE not in flags
-                                and halacha_quality.lexical_near_duplicate(
-                                    h["rule_statement"], neighbor["rule_statement"])):
+                        if (action == "flag"
+                                and halacha_quality.FLAG_NEAR_DUPLICATE not in flags):
                             flags.append(halacha_quality.FLAG_NEAR_DUPLICATE)
 
                 confidence = float(h.get("confidence", 0.0))

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

@@ -244,6 +244,35 @@ def lexical_near_duplicate(
             or normalized_levenshtein(a, b) >= levenshtein_min)
 
 
+def dedup_action(
+    dist: float, rule_new: str, rule_neighbor: str,
+    dedup_distance: float, band_distance: float,
+) -> str:
+    """Decide a fresh halacha's fate vs its nearest same-precedent neighbor (#82.4).
+
+    PAIRWISE by construction — it compares the new rule to exactly ONE neighbor
+    (the nearest already-stored one), never to a cluster, so dedup-on-insert can
+    NEVER collapse a chain A~B~C into a single row even when A and C are
+    distinct: each insert is an independent pairwise decision and only the
+    *incoming* row is ever skipped (no existing row is merged or deleted). This
+    is the over-merge guard (#82.6) — connected-components closure, the central
+    over-merge risk in entity-resolution, is deliberately NOT performed here.
+
+    ``dist`` is cosine distance (1 − cosine sim) to the neighbor. Returns:
+      * 'skip' — semantic duplicate (dist ≤ dedup_distance): drop the incoming
+        row; the caller unions its provenance (cites) into the surviving
+        neighbor rather than blind-dropping it.
+      * 'flag' — lexical tail (dedup_distance < dist ≤ band_distance AND high
+        lexical overlap): keep the row but mark near_duplicate → chair review.
+      * 'keep' — distinct enough: store normally.
+    """
+    if dist <= dedup_distance:
+        return "skip"
+    if dist <= band_distance and lexical_near_duplicate(rule_new, rule_neighbor):
+        return "flag"
+    return "keep"
+
+
 # ── Aggregate ──
 
 FLAG_NON_DECISION = "non_decision"

mcp-server/tests/test_halacha_dedup_action.py

@@ -0,0 +1,53 @@
+"""Tests for #82.4 / #82.6 — dedup-on-insert decision + over-merge guard.
+
+``halacha_quality.dedup_action`` is the PAIRWISE decision a fresh halacha makes
+against its single nearest same-precedent neighbor: skip (semantic dup), flag
+(lexical tail), or keep. It compares to exactly ONE neighbor and only ever drops
+the *incoming* row, so a chain A~B~C can never collapse to one row — the
+over-merge guard (#82.6). Pure/offline.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from legal_mcp.services import halacha_quality as hq
+
+# operating point: DEDUP_COSINE=0.93 → dedup_distance=0.07 ; BAND=0.83 → 0.17
+DEDUP_D = 1.0 - 0.93
+BAND_D = 1.0 - 0.83
+
+SIMILAR_A = "מיצוי הליכים הוא תנאי סף להגשת ערר לוועדה"
+SIMILAR_B = "מיצוי הליכים הוא תנאי סף להגשת הערר לוועדה"
+DIFFERENT = "מתחם שיקול הדעת התכנוני של הוועדה המקומית רחב"
+
+
+def test_skip_below_dedup_distance():
+    # cosine ≥ 0.93 (dist ≤ 0.07) → skip, regardless of wording
+    assert hq.dedup_action(0.03, DIFFERENT, SIMILAR_A, DEDUP_D, BAND_D) == "skip"
+    assert hq.dedup_action(0.05, SIMILAR_A, SIMILAR_B, DEDUP_D, BAND_D) == "skip"
+
+
+def test_flag_in_lexical_tail():
+    # in the 0.07–0.17 band AND lexically near → flag (not skip, not keep)
+    assert hq.dedup_action(0.12, SIMILAR_A, SIMILAR_B, DEDUP_D, BAND_D) == "flag"
+
+
+def test_keep_in_tail_when_not_lexically_similar():
+    # in the band but lexically distinct → keep (don't flag a different rule)
+    assert hq.dedup_action(0.12, DIFFERENT, SIMILAR_A, DEDUP_D, BAND_D) == "keep"
+
+
+def test_over_merge_guard_distinct_rule_kept():
+    """Beyond the band, even a lexically-similar rule is KEPT — and because the
+    decision is pairwise (one neighbor, incoming-only drop), a chain A~B~C with
+    A,C distinct never collapses to a single row (#82.6)."""
+    assert hq.dedup_action(0.30, SIMILAR_A, SIMILAR_B, DEDUP_D, BAND_D) == "keep"
+    assert hq.dedup_action(0.50, DIFFERENT, SIMILAR_A, DEDUP_D, BAND_D) == "keep"
+
+
+def test_boundary_exactly_at_band_edge():
+    # dist == band_distance is still within the tail (≤), lexical → flag
+    assert hq.dedup_action(BAND_D, SIMILAR_A, SIMILAR_B, DEDUP_D, BAND_D) == "flag"
+    # just past the band → keep
+    assert hq.dedup_action(BAND_D + 0.001, SIMILAR_A, SIMILAR_B, DEDUP_D, BAND_D) == "keep"