Proxy-marker weaken-don't-withhold + OCR pre-classification short-circuit

[WilfordGrimley]

Summary

Two independent, pre-197k-remainder correctness fixes (owner-approved), landed as separate commits on this branch so either is separately revertable:

1. fix: proxy-marker weakens no longer withholds a join-key match (MPCAutofill/cardpicker/local_calculate_verdicts.py)

• _apply_agreement_checks no longer returns a proxy-marker-veto skip when legal_line_proxy_marker_detected — a live read-only trace (2026-07-21) found this discarding 1,552 already-validated candidate matches, 99.4% with a real, DB-matching set/number parse.
• Owner ruling (docs/features/catalog-completion-plan.md's "Recovery-arc lessons" item 1's own "CRITICAL correction," encoded verbatim there): the marker is catalog-required on every genuine upload, real printings' proxies included — its presence carries no discriminating power over whether a specific match is right or wrong.
• Confidence-tier choice: no change at all (not a weaker tier). Reasoning: unlike artist-OCR disagreement (a genuine, informative cross-check that covaries with match correctness), a catalog-required field that's true across nearly the whole eligible population is closer to a constant than a signal — inventing a specific weaker number for it would repeat the exact "threshold invented, not measured" mistake this module's own quality/integrity gating note already declines to make for blur_variance/image_entropy. Full reasoning is in the inline comment at the check's former location.
• reparse_collector_evidence gets a new --selector proxy-marker-veto --stage-d-run-id RUN_ID (mirrors --selector no-text structurally) to retract the ~1,552 stale skip rows — immediately actionable (unlike no-text, no re-extraction needed; the code fix alone changes the conclusion for the same stored evidence). Dry-run by default.
• docs/theory.md §7's g4 enumeration and docs/features/catalog-completion-plan.md's "moderator-flag veto"/slow-path sections updated in place to describe the corrected behavior.

2. perf: short-circuit collector_line_ocr escalation on digit-free tier 1 (MPCAutofill/cardpicker/image_evidence.py)

• Implements the merged plan's spec exactly (docs/features/catalog-completion-plan.md's "Recovery-arc lessons" item 1): once both tier-1 attempts fail to parse a collector number, if neither attempt's raw text contains a single digit character, skip tiers 2-3 (6 more tesseract calls) and go straight to no-text — measured at 99.7% of a real no-text cohort, zero-loss on that sample.
• A digit-bearing tier-1 failure still escalates through every tier exactly as before (_contains_digit is a coarser, cheaper check than _COLLECTOR_NUMBER_RE, so this can only ever short-circuit a strict subset of the existing no-text population).
• Escape hatch: compute_card_evidence(..., short_circuit=...) param, run_image_evidence_cohort --no-shortcircuit CLI flag, and STAGE_C_NO_SHORTCIRCUIT env var — all default to short-circuit ON.
• run_image_evidence_cohort now counts and reports short_circuited=N in its progress/DONE summary lines, so the real 197k-card remainder run itself produces the plan's own "open verification gap" measurement data (whether any short-circuited card would have recovered at a later tier).

Test plan

• test_local_calculate_verdicts.py (58 tests): proxy-marker no longer vetoes a direct match, a symbol-tiebroken match; marker present/absent reaches the identical verdict; marker still doesn't affect a genuine no-match; marker no longer masks a real copyright-year mismatch (that withhold now fires instead).
• test_reparse_collector_evidence.py (22 tests): new TestSelectCardIdsProxyMarkerVeto (2 tests) + test_proxy_marker_veto_selector_recovers_immediately_no_reextraction_needed + test_proxy_marker_veto_selector_end_to_end_recasts_the_real_vote (full two-step runbook proof, resolved-consensus safety gate covered by existing shared tests) + CLI --stage-d-run-id requirement test.
• test_image_evidence.py: short-circuit fires by default on a digit-free blank crop (2 tesseract calls, not 8); --short_circuit=False preserves the full 8-attempt escalation; a digit-bearing-but-unparseable tier-1 read still escalates through every tier by default; short-circuit only fires once both tier-1 attempts are confirmed digit-free.
• test_run_image_evidence_cohort.py: _compute_one_card/_run_cohort thread the new short_circuit param and 4th return value through correctly.
• Full backend suite (host venv against live prod containers' exposed Postgres/ES, ephemeral test_* DB, no prod data touched): 1370 passed, 4 skipped (same CI-documented named skips), zero regressions.
• Full pre-commit (ruff/isort/black/mypy/prettier) clean on every changed file.
• No migration involved — nothing to run against the live DB, nothing deferred for that reason.

Runbook additions (for the 197k remainder run)

Re-scan for the ~1,552 stale proxy-marker-veto rows (dry-run first per the binding "mandatory dry-run before any --selector state-clear write" rule):

manage.py reparse_collector_evidence --selector proxy-marker-veto --stage-d-run-id <RUN_ID>
manage.py reparse_collector_evidence --selector proxy-marker-veto --stage-d-run-id <RUN_ID> --write

New flag for the remainder harvest command (default is short-circuit ON — no flag needed for the normal run; only pass this for a measurement/comparison run):

manage.py run_image_evidence_cohort --no-shortcircuit ...

🤖 Generated with Claude Code