Part 4 (LANDS): artist-decomposed identification module + HOLD #B prep

[WilfordGrimley]

Summary

• Builds Part 4 (LANDS) per docs/features/catalog-completion-plan.md's spec: artist-decomposed identification for names whose candidate count blocks the normal phash engine (basic lands + any over-cap name).
• Mirrors Part 3's exact shape: dry_run=True default, run_id, PilotRunLedger lifecycle, verify_zero_resolutions gate after any real write, --write gate on the management command.
• Confidence-tier split (0.85 singleton / 0.8 tiebreak) owner-clarified since the plan doc's "artist+art agree" vs "art-within-artist" phrasing was genuinely ambiguous - full reasoning in the module docstring.
• Caught and fixed two real bugs during review: an artist-extraction-rate denominator bug (was dividing by sampled instead of fetch_attempted, which would have diluted the rate with budget-exhausted skips) and an OCR-confidence recomputation bug (was deriving confidence from .detail truthiness instead of using EngineVote.confidence directly).
• 16 new tests (synthetic fixtures, no network), all passing. Full suite: 862 passed, 130/130 snapshots, only pre-existing known-bucket failures (moxfield x2, test_sources.py fixture-path x2), unrelated to this change.
• ruff/isort/black/mypy/prettier all clean via the real pre-commit hook set.

HOLD #B - still open

The real volume-check numbers (land-pool size, artist-extraction rate on a 300-card sample, per-name candidate counts pre/post artist filter) require running manage.py local_lands_identify (dry-run, the default) against the real production database. This session doesn't have that access (same denial class as earlier Docker/DB access attempts this session) - not re-attempted a fourth way. The command is built, tested, and ready for whoever has that access to run.

Test plan

• pytest cardpicker/tests/test_local_lands_identify.py - 16/16 pass
• pytest cardpicker/tests/ (full suite) - 862 passed, only known-bucket failures
• pre-commit run --files <changed> - ruff/isort/black/mypy/prettier all clean
• HOLD #B's real dry-run against production data - needs DB access this session doesn't have

Generated with Claude Code