#000049 §7 #21: candidate-clause restriction in ShadowNLI.check — helps, doesn't close it

candidate_clauses() — NLI now runs only on the top-N source clauses by
content-token overlap with the answer claim (max_candidate_clauses=6),
not the whole context; records n_candidate_clauses / best_clause_overlap
/ recombination_risk. Synthetic sweep unchanged (28/28 recombination,
0/26 legit FP, mean 1.45 candidate clauses/record). Real-traffic smoke
re-run: STRICT would-demote 30% → 20%, overall 47% → 33% — better, not
fixed; recombination-risk split doesn't separate either. Residual STRICT
false-contras at ~0.83-0.92 → θc would need ≈ 0.90 (vs the clean-set
0.5); at θc=0.90 the data in hand gives 27/28 synthetic recall, 0/26
legit FP, 0/10 smoke STRICT FP — but n=10 is too small to set on.
Next: a fuller ARBORIST_NLI_SHADOW=1 bench-qa run → sweep θc on hundreds
of STRICT cells → confirm → set it. θc stays 0.5; runtime NLI demotion
stays off. Production verifier unchanged; falsification-hard stays 10/12.

tests/test_nli_shadow.py

@@ -16,7 +16,7 @@ from pathlib import Path
 import pytest
 
 from arborist.qa.nli import ShadowNLI, ShadowResult, shadow_check, load_manifest
-from arborist.qa.nli.shadow import clauses, _resolve_label_indices
+from arborist.qa.nli.shadow import clauses, candidate_clauses, _content_tokens, _resolve_label_indices
 
 
 # --- clause splitter -------------------------------------------------------
@@ -42,6 +42,58 @@ def test_clauses_empty_or_blank():
     assert clauses(None) == []  # type: ignore[arg-type]
 
 
+# --- candidate-clause restriction (§7 #5 step 3 / §7 #20) ------------------
+
+def test_content_tokens_filters_short_and_stopwords():
+    toks = _content_tokens("The Mercury is the largest planet in the Solar System.")
+    assert "mercury" in toks and "largest" in toks and "planet" in toks and "solar" in toks
+    assert "the" not in toks and "is" not in toks and "in" not in toks  # stopword / too short
+
+
+def test_candidate_clauses_ranks_by_overlap_and_caps():
+    claim = "Mercury is the largest planet in the Solar System."
+    source = ("Jupiter is the largest planet in the Solar System. "
+              "Mercury is the smallest planet in the Solar System. "
+              "The Roman god Mercury was the messenger of the gods. "
+              "Bananas are yellow.")
+    cand = candidate_clauses(claim, source, k=6)
+    # the banana clause shares no content tokens → dropped
+    assert all("Bananas" not in c for c, _ in cand)
+    # the two planet clauses are the highest-overlap ones
+    assert "Jupiter is the largest planet" in cand[0][0] or "Mercury is the smallest planet" in cand[0][0]
+    assert cand[0][1] > cand[-1][1] or len(cand) == 1
+    # cap respected
+    assert len(candidate_clauses(claim, source, k=1)) == 1
+
+
+def test_candidate_clauses_empty_when_no_overlap():
+    assert candidate_clauses("Bananas are yellow.", "Mercury is the smallest planet.", k=6) == []
+    assert candidate_clauses("", "anything at all here", k=6) == []
+
+
+def test_check_restricts_to_candidate_clauses_not_whole_haystack():
+    # a long source where only 2 clauses touch the claim — the result must
+    # report n_candidate_clauses << n_clauses (this is the §7 #20 fix).
+    haystack = " ".join(f"Unrelated fact number {i} about pottery." for i in range(40))
+    source = "Jupiter is the largest planet. " + haystack + " Mercury is the smallest planet."
+    res = shadow_check("Mercury is the largest planet.", source)
+    assert res.n_clauses >= 40
+    assert res.n_candidate_clauses <= 6
+    assert res.n_candidate_clauses < res.n_clauses
+
+
+def test_check_no_candidate_clauses_is_no_demote():
+    res = shadow_check("Mercury is the largest planet.",
+                       "Bananas are yellow. Pottery is ancient. The weather is fine.")
+    assert res.would_demote is False
+    if res.available:
+        assert res.reason == "no_candidate_clauses" and res.n_candidate_clauses == 0
+
+
+def test_manifest_has_candidate_clause_cap():
+    assert load_manifest()["max_candidate_clauses"] == 6
+
+
 # --- manifest --------------------------------------------------------------
 
 def test_manifest_has_required_fields():