bench/fixtures/5f: harvest 40 corpus-derived falsification fixtures (#000037 → #000025)

Closes the controller → 5F battery loop fox designed: #000037 Phase
1 produces FalsificationFixtureProposal records from high-divergence
providence_cache rows; this commit lands the harvester that turns
those proposals into a real 5F fixture pack the falsification
battery exercises every test run.

bench/scripts/harvest_falsification_proposals.py — reads qa.db,
filters live rows with witness_divergence = (n_unverified / n_quotes)
>= 0.5, stratifies by audit_mode, picks 20 HYBRID + 20 UNGROUNDED
top-by-cache_key for determinism, writes embedded-mode fixtures
to bench/fixtures/5f/falsification-harvested-v1.jsonl.

Each fixture carries `_harvest_meta` with the source cache_key,
divergence at harvest time, audit_mode at harvest time, and the
ticket reference (#000037 §13 step 11). Embedded mode — uses
`observed_violations` directly without calling verify_quotes
again; `answer_text` preserved verbatim for debugging.

Pack composition (initial harvest 2026-05-10):

  20 UNGROUNDED         (expected_reason: UNGROUNDED)
   9 HYBRID_QUOTE       (NEW motif — not in falsification-v1.jsonl)
   7 HYBRID_CLAIM_LATTICE (NEW motif — not in falsification-v1.jsonl)
   4 HYBRID_PARAPHRASE  (already covered in v1)

Two harness tests pin the pack:
- test_5f_falsification_harvested_pack_runs_clean: 100% pass-rate,
  every fixture carries traceable _harvest_meta.
- test_5f_falsification_harvested_pack_widens_motif_coverage: the
  HYBRID_QUOTE / HYBRID_CLAIM_LATTICE motifs surface from real
  corpus (loud-fail if harvest rotation drops them).

Makefile: `make bench-5f-harvest` re-runs the harvester. Parameters
exposed: HARVEST_QA_DB, HARVEST_OUT, HARVEST_THRESHOLD,
HARVEST_SAMPLE_PER_BUCKET.

Full suite: 2328 passed, 37 skipped (+2 from the two new pins).

tests/test_bench_batteries.py

@@ -671,6 +671,57 @@ def test_5f_falsification_covers_every_documented_motif():
     )
 
 
+def test_5f_falsification_harvested_pack_runs_clean():
+    """#000037 Phase 1 → 5F loop: the harvested pack at
+    falsification-harvested-v1.jsonl was generated by
+    bench/scripts/harvest_falsification_proposals.py from real qa.db
+    high-divergence rows. Pin: pack runs at 100% pass-rate, all
+    fixtures route through the embedded runner path, every fixture
+    carries the harvest-attribution metadata."""
+    import json
+
+    fixture_path = F5F / "falsification-harvested-v1.jsonl"
+    res = b_5f.run_falsification(fixture_path)
+    assert res.fail_count == 0
+    assert res.metrics["error_detection_rate"] == 1.0
+    # Every fixture must carry the harvest metadata so future
+    # regenerations are traceable to the source providence_cache
+    # row.
+    with fixture_path.open() as f:
+        next(f)  # _meta header
+        for line in f:
+            row = json.loads(line)
+            meta = row.get("_harvest_meta", {})
+            assert meta.get("source_ticket") == "#000037 §13 step 11"
+            assert meta.get("harvested_from") == "providence_cache"
+            assert isinstance(meta.get("witness_divergence"), (int, float))
+            assert meta["witness_divergence"] >= 0.5
+            assert meta.get("audit_mode_at_harvest") in ("HYBRID", "UNGROUNDED")
+
+
+def test_5f_falsification_harvested_pack_widens_motif_coverage():
+    """The harvested pack introduces real-corpus motif tags that the
+    hand-curated falsification-v1.jsonl doesn't have. Pin: the
+    HYBRID_QUOTE + HYBRID_CLAIM_LATTICE pair appear in the harvested
+    pack — concrete evidence that the harvest is doing widening
+    work, not just duplicating the curated set."""
+    import json
+
+    fixture_path = F5F / "falsification-harvested-v1.jsonl"
+    motifs: set[str] = set()
+    with fixture_path.open() as f:
+        for line in f:
+            d = json.loads(line)
+            if "expected_reason" in d:
+                motifs.add(d["expected_reason"])
+    # Harvest contains HYBRID variants surfaced from corpus that the
+    # hand-curated set lacks. If the harvested DB drifts and these
+    # motifs disappear, the test loud-fails — by then the harvest
+    # rotation has lost signal and the pack should be regenerated.
+    assert "HYBRID_QUOTE" in motifs or "HYBRID_CLAIM_LATTICE" in motifs
+    assert "UNGROUNDED" in motifs
+
+
 def test_5f_falsification_live_helper_calls_real_verifier():
     """UNGROUNDED + STRICT_SPAN signals come straight from verify_quotes."""
     from bench.batteries.b_5f import _live_falsification_violations