qa/live: lock 3 bench standouts as STRICT-rate regression gates

Bench finding (2026-04-30, n=3 across 24 questions): JSON mode hit
S:3 H:0 U:0 on three questions where pointer mode landed S:0 H:3 U:0
or quote mode landed S:0 H:0 U:3 — clean modal divergence. Pinning
these as live fixtures gives us regression detection: a JSON-mode
quality drop on these questions reverts the STRICT signal first.

  test_mona_lisa_strict_and_names_leonardo  — was _grounded, tightened
    to STRICT (bench: quote 0S, pointer 0S, JSON 3S)
  test_supermans_girlfriend_strict_and_names_lois_lane (NEW)
    — relationship-shape, JSON's strong suit (quote 0S, pointer 0S,
    JSON 3S)
  test_cold_fusion_breakthrough_year_strict_and_names_1989 (NEW)
    — out-of-corpus rephrasing, atomic-claim rule unlocked grounding
    (quote 0S 3U, pointer 3S, JSON 3S)

Per fox's "scoreboard + gates" docstring: bench gives aggregate
counts; these fixtures fail by name when the scoreboard moves.

tests/test_qa_quality_live.py

@@ -124,13 +124,49 @@ def test_linux_kernel_grounded_and_names_torvalds():
         f"answer missing 'Torvalds': {_answer_lower(r)[:200]}"
 
 
-def test_mona_lisa_grounded_and_names_leonardo():
+def test_mona_lisa_strict_and_names_leonardo():
+    """Bench standout (2026-04-30, n=3): JSON mode hit S:3 H:0 U:0
+    while quote landed 0S and pointer 0S. Tightened from _grounded
+    to STRICT to lock in the JSON-mode unique win as a regression
+    gate — drop in JSON-mode quality on this question reverts the
+    STRICT signal first."""
     r = _ask("who painted the mona lisa?")
-    assert _grounded(r), f"audit_mode={r.get('audit_mode')!r}"
+    assert r.get("audit_mode") == "STRICT", \
+        f"expected STRICT (bench 3/3 STRICT in JSON), got {r.get('audit_mode')}"
     assert "leonardo" in _answer_lower(r), \
         f"answer missing 'Leonardo': {_answer_lower(r)[:200]}"
 
 
+def test_supermans_girlfriend_strict_and_names_lois_lane():
+    """Bench standout (2026-04-30, n=3): JSON mode hit S:3 H:0 U:0
+    while pointer landed S:0 H:3 (every grounded answer demoted by
+    lazy-anchor or pointer-saturation). The relationship-shape
+    question 'who is X's girlfriend' is JSON's strong suit — atomic
+    claims, single-name target. Gate: STRICT + 'lois' or 'lane'."""
+    r = _ask("who is supermans girlfriend?")
+    assert r.get("audit_mode") == "STRICT", \
+        f"expected STRICT (bench 3/3 STRICT in JSON), got {r.get('audit_mode')}"
+    txt = _answer_lower(r)
+    assert "lois" in txt or "lane" in txt, \
+        f"answer missing Lois/Lane: {txt[:200]}"
+
+
+def test_cold_fusion_breakthrough_year_strict_and_names_1989():
+    """Bench standout (2026-04-30, n=3): originally classified
+    out-of-corpus but JSON mode hit S:3 H:0 U:0 — the corpus has
+    Pons & Fleischmann's 1989 announcement covered. Quote mode
+    landed 0S 0H 3U (UNGROUNDED) so this fixture also catches
+    quote/JSON divergence. Gate: STRICT + a 1989-era marker
+    (year, Pons, Fleischmann, or Utah)."""
+    r = _ask("what year does our cold fusion breakthrough happen?")
+    assert r.get("audit_mode") == "STRICT", \
+        f"expected STRICT (bench 3/3 STRICT in JSON), got {r.get('audit_mode')}"
+    txt = _answer_lower(r)
+    cold_fusion_markers = ("1989", "pons", "fleischmann", "utah")
+    assert any(m in txt for m in cold_fusion_markers), \
+        f"answer missing 1989-era cold-fusion markers: {txt[:300]}"
+
+
 # ---------------------------------------------------------------- retrieval disambiguation