Stage D: gate the frame agreement veto on artist_ocr having actually run

[WilfordGrimley]

Item 4 of the "essential major blockers" brief, narrowed per its own instruction: "If the full sweep is too large, fix the artist_ocr → frame-veto path alone — that is the one producing permanent wrong answers."

The defect

run_join_key_calculator filters eligibility on extractor_versions__has_key="collector_line_ocr" and then reads six extractors' fields ungated: layout_class, illus_anchor_fired, legal_line_*, artist_ocr_name, symbol_phash, image_is_truncated.

Five degrade permissively — a blank legal line reads as "nothing to compare", a null image_is_truncated reads as "not truncated". A permissive degradation is recoverable, because the human-backed consensus gate still stands between it and any resolution.

The frame agreement veto degrades STRICT, and that is why it is the one fixed. ImageEvidence.illus_anchor_fired is nullable, so bool(None) is False — indistinguishable from "artist_ocr ran and found no anchor". On a card with no collector number either, classify_frame_style returns "modern" for a card it has no anchor evidence about at all, and a genuine old-frame printing is vetoed frame-mismatch.

That reason is deliberately not in JOIN_KEY_RESCANNABLE_SKIP_REASONS, so the wrong conclusion is permanent for that content hash: the card never becomes eligible again, and no later Stage C pass can undo it. A wrong answer nothing can revisit is strictly worse than a missing one.

The fix

FRAME_CHECK_REQUIRED_EXTRACTOR_KEYS = frozenset({"collector_line_ocr", "artist_ocr"}) — both extractors classify_frame_style actually reads, since its two arguments come from two different ones and the eligibility query only ever guaranteed the first.

An absent key skips the check, not the card. The match stands at its already-computed confidence, and once Stage C fills artist_ocr in, the check runs for real. This is not a new rule — it is the "missing data is not evidence" rule the function's own docstring already states, and which the copyright-year check and the metadata is None case already follow.

Declared per-CHECK, not per-calculator. local_detect_ai_art, local_lands_identify and local_layout_class_cast each declare one module-level REQUIRED_EXTRACTOR_KEYS because each runs one check. This calculator runs several with genuinely different needs — the join-key deduction itself needs only the collector line — so a calculator-wide gate would drop cards that have everything their own decision requires.

artist_ocr is at 220,579/220,579 coverage in production, so this changes nothing about the current catalogue. It removes a trap; it does not loosen a live gate.

Not in scope

The other five ungated reads are left as-is and named in the code comment. They degrade permissively, which is recoverable, and closing them is a coverage/behaviour tradeoff worth measuring on its own rather than folding into a permanent-wrong-answer fix.

Verification

mutation	expected red	result
remove the gate (restore the ungated read)	old-frame card wrongly vetoed	1 failed, 179 passed
shrink the gate to collector_line_ocr alone	the defect returns + the constant tripwire	2 failed, 178 passed
over-apply — skip the frame check unconditionally	pre-existing test_frame_mismatch_withholds_the_match + the new control	2 failed, 178 passed
restored	—	180 passed

The third mutation is the important one: it proves the fix is not a licence to drop the veto. The control test shows the same card, same evidence, one extra manifest key is still vetoed.

• pytest cardpicker/tests/ — 3464 passed, 11 skipped (full suite)
• docs_lint.py --strict — clean
• pre-commit (ruff/isort/black/mypy/prettier) — clean
• No migration.

The shared _evidence test fixture now carries both OCR extractors by default, matching production. Without that, every existing frame-veto test would have become a silent no-op under the new gate — worth flagging explicitly, since a fixture that quietly disarms the tests around it is the same class of problem as the ungated read itself.

Docs

docs/identification-pipeline.md g4 — new "A check whose extractor never ran does not run" paragraph, stating the permissive-vs-strict distinction and why only one of the six needed gating (living wiki doc; no dated report).

🤖 Generated with Claude Code

https://claude.ai/code/session_013NhYmT1PxCcyemA16dFDxN