feat(halacha): NLI entailment validator via claude_session (#81.3) + task #86

#81.3 — a post-extraction validator that flags halachot whose rule_statement is
NOT entailed by its supporting_quote (the model over-reaching beyond its source).

- Engine: claude_session-as-judge (local CLI, zero API cost) per chaim's standing
  preference — one batched judge call per chunk, NOT a hosted NLI model.
- Pure, unit-tested helpers in halacha_quality: NLI_SYSTEM, build_nli_prompt,
  parse_nli_verdicts (fails OPEN — any shape/label ambiguity → 'entailed').
- halacha_extractor._nli_check wraps the call; fails OPEN on any error (e.g. no
  CLI in the container) so a flaky judge never blocks a genuine halacha.
- Non-entailed (neutral/contradiction) → quality_flag 'nli_unsupported' which
  blocks auto-approve (routes to pending_review) via the existing store gate.
- config: HALACHA_NLI_ENABLED/MODEL/EFFORT (effort 'low' — entailment is simple).

Verified: suite 166 passed (10 new); LIVE smoke test against the real claude CLI
returned ['entailed','neutral'] for a supported vs unsupported rule.

Also commits TaskMaster #86 (Nevo preamble/ratio: anti-contamination strip fix +
gold-set benchmark) capturing today's strip_nevo_preamble findings.

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

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

@@ -134,6 +134,55 @@ 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)
+
+
+# ── NLI entailment check (rule_statement ⊨ supporting_quote) — #81.3 ──
+#
+# Pure prompt-builder + verdict-parser; the LLM call itself runs through
+# claude_session in halacha_extractor (local CLI, zero cost). A rule that the
+# quote does not actually support (neutral) or contradicts is the model
+# over-reaching beyond its source — flag it (blocks auto-approve). EVERYTHING
+# here fails OPEN: any parse ambiguity resolves to "entailed" so a flaky judge
+# never blocks a genuine halacha.
+
+NLI_SYSTEM = (
+    "אתה בודק היסק (entailment) משפטי. לכל זוג {כלל, ציטוט} החלט האם **הכלל נובע מהציטוט** — "
+    "כלומר הציטוט תומך בכלל ואינו מרחיב מעבר למה שנכתב בו. שלוש תוויות בלבד:\n"
+    "- entailed = הכלל נתמך במלואו בציטוט.\n"
+    "- neutral = הציטוט אינו תומך בכלל (הכלל מרחיב/מוסיף מעבר לציטוט).\n"
+    "- contradiction = הכלל סותר את הציטוט.\n"
+    'החזר JSON array בלבד באורך מספר הזוגות, לדוגמה: ["entailed","neutral",...]. '
+    "ללא markdown, ללא הסבר."
+)
+
+_NLI_LABELS = {"entailed", "neutral", "contradiction"}
+
+
+def build_nli_prompt(items: list[dict]) -> str:
+    """Build the user message: a numbered list of {rule, quote} pairs."""
+    blocks = []
+    for i, h in enumerate(items, 1):
+        rule = (h.get("rule_statement") or "").strip()
+        quote = (h.get("supporting_quote") or "").strip()
+        blocks.append(f"### זוג {i}\nכלל: {rule}\nציטוט: {quote}")
+    return "\n\n".join(blocks)
+
+
+def parse_nli_verdicts(raw, n: int) -> list[str]:
+    """Coerce the judge's output into exactly ``n`` labels — fail-open.
+
+    Any shape mismatch / unknown label resolves to 'entailed' so a flaky or
+    unavailable judge never blocks a halacha.
+    """
+    if not isinstance(raw, list) or len(raw) != n:
+        return ["entailed"] * n
+    out: list[str] = []
+    for item in raw:
+        v = item.get("verdict") if isinstance(item, dict) else item
+        v = str(v or "").strip().lower()
+        out.append(v if v in _NLI_LABELS else "entailed")
+    return out
 
 
 def compute_quality_flags(