Merge pull request 'fix(principles): cull resumability + cap criterion-A (consensus-first) (#152)' (#305) from worktree-canonical-synthesis into main

Merge PR #305: cull resumability + cap criterion-A (#152)

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

@@ -136,7 +136,10 @@ def apply_cap(judged: list[dict], max_new: int | None = None) -> list[dict]:
     """
     max_new = config.HALACHA_PANEL_MAX_NEW if max_new is None else max_new
     survivors = [j for j in judged if j.get("verdict") in ("approved", "pending_review")]
-    survivors.sort(key=lambda j: j.get("score", 0.0), reverse=True)
+    # Rank consensus-first, then confidence (chaim 2026-06-20): a unanimous 3-vote
+    # principle outranks a 2-vote one regardless of score — cross-model agreement is
+    # the more reliable keep signal (AC1=0.92). Mirrors cluster_candidates' ordering.
+    survivors.sort(key=lambda j: (j.get("votes", 0), j.get("score", 0.0)), reverse=True)
     keep_ids = {id(j) for j in survivors[:max_new]}
     out = []
     for j in judged:

mcp-server/tests/test_panel_extraction.py

@@ -106,17 +106,27 @@ def test_cluster_same_model_twice_counts_one_vote_keeps_best_score():
     assert cl["rule_statement"] == "X"
 
 
-def test_apply_cap_downgrades_over_cap_survivors_by_score():
+def test_apply_cap_downgrades_over_cap_survivors_by_votes_then_score():
     judged = [
-        {"verdict": "approved", "score": 0.9},
-        {"verdict": "approved", "score": 0.7},
-        {"verdict": "pending_review", "score": 0.8},
-        {"verdict": "rejected", "score": 0.95},   # already rejected stays
+        {"verdict": "approved", "votes": 3, "score": 0.9},
+        {"verdict": "approved", "votes": 3, "score": 0.7},
+        {"verdict": "pending_review", "votes": 2, "score": 0.8},
+        {"verdict": "rejected", "votes": 1, "score": 0.95},   # already rejected stays
     ]
     out = pe.apply_cap(judged, max_new=2)
     fv = [j["final_verdict"] for j in out]
-    # top-2 survivors by score = 0.9(approved) + 0.8(pending); 0.7 → over cap → rejected
-    assert fv == ["approved", "rejected", "pending_review", "rejected"]
+    # top-2 by (votes,score) = both 3-vote (0.9, 0.7); the 2-vote/0.8 → over cap → rejected
+    assert fv == ["approved", "approved", "rejected", "rejected"]
+
+
+def test_apply_cap_votes_outrank_score():
+    # a 2-vote/0.95 must NOT beat a 3-vote/0.80 — consensus dominates confidence
+    judged = [
+        {"verdict": "approved", "votes": 2, "score": 0.95},
+        {"verdict": "approved", "votes": 3, "score": 0.80},
+    ]
+    out = pe.apply_cap(judged, max_new=1)
+    assert [j["final_verdict"] for j in out] == ["rejected", "approved"]
 
 
 def test_apply_cap_keeps_all_when_under_cap():

scripts/cull_principles.py

@@ -50,11 +50,16 @@ _JUDGE_CONCURRENCY = 4
 async def _decisions(limit, sample):
     """case_law ids that have 'original' principles, with source metadata."""
     pool = await db.get_pool()
+    # Resumable: skip decisions already culled (their survivors carry reviewer
+    # 'cull:%'), so a continuation run after a ceiling-stop never re-judges them.
+    # (Dry-run never sets that reviewer, so nothing is skipped in dry-run.)
     rows = await pool.fetch(
         "SELECT cl.id, cl.case_number, cl.source_kind, cl.is_binding, "
         "       count(*) AS n "
         "FROM halachot h JOIN case_law cl ON cl.id = h.case_law_id "
         "WHERE h.instance_type = 'original' AND h.review_status <> 'rejected' "
+        "AND NOT EXISTS (SELECT 1 FROM halachot h2 WHERE h2.case_law_id = cl.id "
+        "  AND h2.instance_type = 'original' AND h2.reviewer LIKE 'cull:%') "
         "GROUP BY cl.id, cl.case_number, cl.source_kind, cl.is_binding "
         "ORDER BY n DESC",
     )