feat(artbox-exemplar-index): self-referential artbox-phash exemplar index (issue #508 phase 1)

[WilfordGrimley]

Summary

1. Adds ArtboxPhashExemplar (+ ArtboxPhashExemplarSeedKind): a labelled association from a card's own current artbox_phash (issue Artbox perceptual-hash extractor: evidence-only, rides the next whole-catalog pass #480's evidence-only extractor) to the illustration_id of the printing that scan was identified as. Sourced exclusively from our own DB, never Scryfall images (PR feat(local_phash): local-only resumable backfill for CanonicalCard.image_hash #694 was closed and deferred to Deferred: cross-source phash reference corpus backfill (PR #694) — supplanted by #508's self-referential exemplar index #697 for exactly this reason — phash comparability needs identical crop geometry, which only our own extractor guarantees).
2. Adds cardpicker/artbox_exemplar_backfill.py: the seeding logic, reusing the batched/resumable/idempotent shape from PR feat(local_phash): local-only resumable backfill for CanonicalCard.image_hash #694 (feat/local-backfill-canonical-hash) — only the source changes, from Scryfall CDN fetches to a scan of our own ImageEvidence. Two seed sources, per the owner's 2026-08-05 extension of Artbox-phash illustration matching: resolution-seeded exemplar index (second deduction path) #508's original human-only design:
   • Human-backed printing resolutions (printing_tag_status == RESOLVED — resolution always requires a human-backed vote per resolve_weighted_consensus's gate, so no per-vote inspection is needed).
   • Unopposed join-key machine votes at 0.75/0.85 confidence (JOIN_KEY_SEED_CONFIDENCE_FLOOR). The artist-disagreement tier (0.65, 38 votes) and the no-match tier (0.6, 16,537 votes — not identifications at all) are both excluded; see that constant's own comment for the full measured distribution and reasoning.
3. Every exemplar records its own seed kind, confidence, and a seed_group_key (mirroring printing_consensus.md5_group_key for the human-resolution case, or the source vote's own pk for the machine case) — a bad seed retracts together with everything it seeded via one filtered delete on that key. Human and machine seeds stay permanently distinguishable (is_human_backed, enforced by a CheckConstraint against seed_kind).
4. Adds backfill_artbox_phash_exemplars (seeding, --dry-run supported) and retract_artbox_phash_exemplars (retraction by seed_group_key/card/illustration/run/source-vote, dry-run by default) management commands.
5. Adds the coverage measurement (measure_unresolved_coverage): of every currently-UNRESOLVED card carrying a current artbox_phash, how many match the exemplar index at d=0 and separately at d≤2 — the honest phase-2 estimate, computed read-only every invocation.
6. Documents the mechanism in docs/identification-pipeline.md's "Parallel detectors" section (the neighbouring Artbox perceptual-hash extractor: evidence-only, rides the next whole-catalog pass #480/OCR-free-path section), stating explicitly what phase 1 does NOT do: no matching calculator, no vote, no consensus change, no touch to the human-backed resolution gate.
7. Single-leaf migration 0102_artbox_phash_exemplar.py on top of 0101_delete_printingtagvote.

Coverage measurement (real, read-only, against the live catalogue)

Executed via a read-only script (mirroring this branch's exact query logic verbatim) inside the live mpcautofill_django container — see DEVIATIONS in the session report for why the literal management command wasn't invoked against production directly. Results, 2026-08-05:

• human_backed_seeded = 12, machine_seeded = 42,050 (29 skipped for stale/missing evidence) → 42,062 exemplars total, covering 17,304 distinct illustration_ids.
• Of 151,100 currently-UNRESOLVED cards carrying a current artbox_phash: 43,400 match at d=0 (28.7%), 46,457 match at d≤2 (30.7%).

This is a strong phase-2 signal — roughly 3 in 10 unresolved cards already have a perceptually-identical or near-identical exemplar in the index, entirely from resolutions and join-key votes that already exist today.

Test plan

• pytest cardpicker/tests/test_artbox_exemplar_backfill.py — 33 tests, all passing (model constraints, both seed passes, confidence floor, staleness/no-match exclusion, idempotent resume, dry-run, retraction, coverage matching).
• Full backend suite (pytest cardpicker/) — 3639 passed, 8 skipped (pre-existing, unrelated), 0 failed.
• pre-commit run on all touched/new files — ruff, isort, black, mypy, prettier all pass.
• makemigrations --check --dry-run — no changes detected (migration matches models exactly); confirmed single leaf on top of 0101_delete_printingtagvote.
• Coverage measurement executed read-only against the live database (see above).