feat(printing-metadata): persist Scryfall layout, measure sideways cohort (issue #693)

[WilfordGrimley]

Summary

• Step 1: adds CanonicalPrintingMetadata.layout (migration 0103_canonicalprintingmetadata_layout), populated verbatim from printing_metadata_import.PrintingMetadataRow.layout — a value that was already parsed but previously discarded once checked against DOUBLE_FACED_LAYOUTS. Follows the model's own established field conventions exactly (CharField(blank=True) like border_color/frame, added to _METADATA_SYNC_FIELDS, wired into the CanonicalPrintingMetadata(...) constructor call). DOUBLE_FACED_LAYOUTS's existing use of row.layout is unchanged.
• Step 2 (redone, same day): the first measurement (41 of 230,488 catalogued cards, 0.018%, 2.4% collector-line parse failure) was retracted — it counted only cards already carrying a canonical_card link, and that link is acquired by having been successfully identified, so a sideways card whose crop lands on rotated content and fails identification could never appear in that cohort. The cohort was drawn from a population selected for not having failed.
• Redo: cohort defined by expected layout instead — Card.name (source metadata, independent of identification) name-matched against CanonicalCard rows whose Scryfall layout is planar/scheme, via the same to_searchable join deductive_backfill.CanonicalNameIndex already uses. Result: 1,093 cards (0.474% of the catalogue), fully containing the old 41-card cohort; only 43 of the 1,093 (3.9%) carry a canonical_card link, confirming the bias directly. Its collector-line parse failure rate is 53.3% (583/1,093) against the correct comparison group — the unlinked population generally, 40.4% (85,217/210,832) — not the whole-catalogue figure (38.4%), which mixes the unlinked population with the already-successfully-identified linked population (17.1%) and understates the real unlinked baseline. That is a genuine 13-point elevated signal the biased denominator hid.
• Rotation trial (the test that elevated signal was supposed to gate): 30 real images sampled live from this cohort's never-linked pool (1,046 cards), fetched via image_cdn_fetch.fetch_card_image, run through the existing collector-line and Illus.-anchor crops unmodified, then again with the whole image rotated 90° clockwise and 90° counter-clockwise. Collector-line OCR produced a plausible-looking parse on 17/30 baseline, 9/30 CW, 7/30 CCW — but validated against each card's own real candidates (local_ocr.validate_against_candidates, the same check Stage D applies), zero were genuine matches in any orientation. The Illus.-anchor artist-line crop found zero in all three orientations too.
• Gate result: crop-box rotation (local_fallback.normalize_crop_box) is not warranted and was not built — not because the cohort is negligible (it isn't, 0.474%) or the failure rate is unremarkable (it isn't, +13 points over baseline), but because the rotation trial found nothing to recover. Matches issue Trimmed-bleed cohort fails collector-line OCR at 12x the bleed rate (remap or classifier tolerance) #683's own finding for the heavily-overlapping trimmed cohort (24/30 of this sample is bleed_class=trimmed): these are disproportionately fan-made custom renders that plausibly never had a machine-readable collector line at all — a content problem, not a crop-geometry problem.
• golden_set.py's 30 pinned cards contain zero sideways-layout cards, so the golden set cannot catch a regression in this path.
• Documents the field itself, the retracted first measurement, the redo methodology, and the rotation trial in docs/features/catalog-completion-plan.md's Stage C section.
• Documents a fresh-worktree makemigrations blocker (missing staticfiles manifest, unrelated to migrations themselves) in docs/troubleshooting.md, hit and resolved while verifying the migration locally.

Test plan

• pytest cardpicker/tests/test_printing_metadata_import.py cardpicker/tests/test_golden_set.py cardpicker/tests/test_local_fallback.py cardpicker/tests/test_image_evidence.py cardpicker/tests/test_run_image_evidence_cohort.py — 307 passed, 0 failed.
• makemigrations --check --dry-run — no pending model changes (migration is complete and single-leaf: 0103 depends on 0102_artbox_phash_exemplar, nothing else branches off 0102).
• pre-commit run (ruff, isort, black, mypy, prettier) green on all touched files.
• docs-lint clean.
• Step 2's redone measurement ran read-only against the live database and its Scryfall bulk cache (migration 0103 is not yet applied to production, so layout was sourced from the bulk-cache file directly, same as the original measurement — not from the not-yet-migrated DB column).
• Rotation trial: 30-card real-image sample, before/after collector-line and artist-line parse outcomes under baseline + 2 rotation directions, validated against real candidates — negative result, documented above.
• CI not watched by this session.