feat(golden-set): pin genuine trimmed real-printing and sideways-layout cards (issue #683 follow-up)

[WilfordGrimley]

What

Closes two known golden-set coverage gaps (cardpicker/golden_set.py), additively — the original 30 pinned cards and their expectations are untouched.

What was measured

Trimmed geometry. Issue #689's "the trimmed cohort is dominated by fan-made custom renders" finding was based on the 8 highest-volume trimmed sources (88.4% of the 2,786-card trimmed population). This PR sampled 7 real cards from 6 previously-unchecked long-tail sources (Monza81, Mursu, MrTeferi, Mithrane, CompC, TorskeRavn, Kevin Kneupper) and visually inspected each fetched image against the live Scryfall API.

Result: the population is not entirely custom renders. 3 of 7 sampled cards carried real, Scryfall-verifiable collector-line data. One of those — a "MH3-AI • EN Kevin Kneupper ft. Midjourney" credit line on "Witch-Blessed Meadow" — looked equally plausible but was confirmed fabricated: the real MH3 #239 printing is by Tyler Walpole (painted art), and "MH3-AI" does not exist as a Scryfall set. Not pinned.

The genuine one, card 35449 ("Fierce Guardianship"), verified against Scryfall's own c20/35 record (artist Randy Vargas, exact flavor-text match), is now pinned. It is a real printing but not a photographic scan — its dimensions match Scryfall's own border_crop image asset, not a camera/scanner capture of a physical card. No genuine photographic bleed-trimmed scan was found in this catalogue.

Sideways layout. Pinned card 130028 ("Bowie Base One"), a real Scryfall layout:planar card (verified via a live layout:planar or layout:scheme Scryfall search, not guessed from the name). Its real extraction run does not show the clean across-the-board abstention a naive prediction would expect: collector_line_ocr/legal_line confidently misparse the card's own sideways credit-line text into a false-positive (set_code="not", collector_number="2024") pair, while artist_ocr correctly abstains despite reading real (but irrelevant) rules-text noise from the same crop region.

What's covered / not covered

All 11 GOLDEN_EXPECTATIONS keys extended for both new ids, each with an inline comment explaining why the card is pinned and what regression it protects against — see golden_set.py's own GOLDEN_CARD_IDS comment for the full verification writeup.

Remains uncovered: a genuine photographic bleed-trimmed scan of a real card (none turned up in this sample — see the trimmed-cohort comment for what that does and doesn't mean). Sideways coverage is a single card from one cohort (planar); scheme layout is not separately represented.

Reconcile with master (2026-08-05)

Master moved to 17345a48 while this PR was open — two siblings landed: #696 (persisted CanonicalPrintingMetadata.layout) and #702 (collapsed the Stage C OCR attempt ladder from 3 tiers to 2; bumped COLLECTOR_LINE_OCR_EXTRACTOR_VERSION/COLLECTOR_LINE_TSV_EXTRACTOR_VERSION v2→v3 and ARTIST_OCR_EXTRACTOR_VERSION v3→v4). Merged (not rebased, to avoid a force-push) at 7757257b.

Conflicts resolved in golden_set.py and docs/troubleshooting.md — both sides were additive in each case, so both bodies of content were kept: master's collector_line_ocr re-verification comment for the original 30 alongside this branch's two new pins and their provenance comment; this branch's rewritten trimmed-cohort closing sentence followed by master's new collectstatic/staticfiles-manifest troubleshooting entry. docs/features/catalog-completion-plan.md auto-merged cleanly (non-overlapping edits); one sentence was added to its golden-set bullet pointing at the open item below.

Known gap, not fully closed by this reconcile: 35449/130028's OCR-derived expectations were recorded before #702's ladder collapse. #702's own re-verification pass covered only the original 30 cards — these two weren't in GOLDEN_CARD_IDS yet when it ran. Breaking down what's actually known:

• Provably unaffected, no live run needed: 35449's collector_line_ocr (already a full abstention under the pre-perf(image-evidence): collapse Stage C OCR attempt ladder from 3 tiers to 2 (issue #677) (re-land of #698) #702 8-attempt ladder — removing an attempt tier can only remove an opportunity to succeed, never create one) and both cards' artist_ocr (same argument, also an abstention).
• Not provably safe, and not independently re-verified against current code: 130028's collector_line_ocr match ({"set_code": "not", "collector_number": "2024"}) and both cards' collector_line_tsv=True — these resolved at some attempt tier under the old ladder, and which tier is unknown without tracing a live run.
• This reconcile could not reach production data to run that check from its own environment: no MPCAutofill/.env present in the worktree (not recreated — out of scope to source one from elsewhere on the box), no production-DB-default Postgres port reachable on the host, and mpcautofill_django (which does have DB access) was off limits to exec/cp under an open ruling for this task's duration. See golden_set.py's own collector_line_ocr comment for the full reasoning trail.
• Open item: someone with production DB access should run a real, no-persistence extract_card_evidence() pass over cards 35449 and 130028 against current (post-perf(image-evidence): collapse Stage C OCR attempt ladder from 3 tiers to 2 (issue #677) (re-land of #698) #702) code and confirm or correct GOLDEN_EXPECTATIONS["collector_line_ocr"][130028] and GOLDEN_EXPECTATIONS["collector_line_tsv"][35449]/[130028].

Test plan

• cardpicker/tests/test_golden_set.py alone — 29 passed
• cardpicker/tests/test_image_evidence.py alone — 115 passed
• Both files together, host venv (order-dependence check per issue Test suite is order-dependent: leaked fetch-failure window trips the envelope across files (8 failures on master) #679) — 144 passed
• .github/scripts/check_extractor_ownership_totality.py — clean (19 declared contributors)
• .github/scripts/check_extractor_manifest_sync.py — clean (11 manifest keys)
• .github/scripts/tests/test_check_extractor_ownership_totality.py — 18 passed
• manage.py makemigrations --check --dry-run — no changes detected
• black/ruff/isort/mypy/prettier — all passed via pre-commit on the merge commit
• Live re-run of extract_card_evidence() for cards 35449/130028 against current (perf(image-evidence): collapse Stage C OCR attempt ladder from 3 tiers to 2 (issue #677) (re-land of #698) #702) code — blocked in this task's environment, see "Reconcile with master" above; tracked as the one open item

Docs

docs/features/catalog-completion-plan.md's golden-set substrate paragraph and docs/troubleshooting.md's issue #683 entry both updated in place to reflect the 32-card total and the refined finding; the golden-set bullet additionally notes the open re-verification item above.