retune(art-edge): within-image relative colour comparison; snip does not clear the validation bar, not persisted

[WilfordGrimley]

Description

Retunes local_art_edge.classify_art_edge_continuity from an ABSOLUTE per-band
pixel-variance test (local_fallback._BORDER_UNIFORMITY_STD_THRESHOLD, tuned
for classify_border_color's different question) to a WITHIN-IMAGE relative
colour comparison: does the band beside the art crop match the border colour
this same image's layout_class (classify_border_color's own output)
already says it has?

This fixes the specific, confirmed defect: dark, textured, off-hue artwork
beside a real border used to read extended purely because it was dark and
flat (low absolute variance), not because it matched or mismatched the
border's actual colour. See
docs/reports/2026-08-06-art-edge-relative-comparison.md
for the full defect reproduction, the retune's own colour-distance/threshold
justification, and the validation methodology.

Validated against Scryfall's own images before any persistence decision
(three cohorts: 30 confirmed extended-art, 20 confirmed borderless, 20
confirmed ordinary framed — label and pixels from the same authoritative
source, unlike the earlier catalogue-cohort measurement this supersedes,
which name-matched uploads to printings without ever confirming the pixels
themselves). Result: does not clear the stated bar — 0/30 recall on
genuine extended-art images (worse than the pre-retune classifier's 1/30 on
the same images), 0 false positives on both negative cohorts. Root cause
traced and documented in the report: classify_border_color's own
"not uniform → borderless" catch-all misreads 20/30 extended-art images, and
for the remaining 10 the trimmed-image edge-band geometry shrinks to ~2px on
a Scryfall image (no bleed margin) — too thin for a reliable colour
reference. Both causes are upstream of this module's scope
(local_fallback.py, PROTECTED CORE).

Per the brief's own stated fallback for a classifier that does not clear
its validation bar: no ImageEvidence.art_edge_class field, no migration, no
Stage C extraction wiring, no EXTRACTOR_OWNERSHIP/manifest entry, no vote
wiring. classify_art_edge_continuity ships as the new implementation,
evidence-only, called from nowhere but its own tests — exactly the state the
pre-retune code was already in, on more defensible internals. Docs updated
to record the measured result (docs/identification-pipeline.md,
docs/pipeline-fidelity-gate.md's calculator roster entry) rather than
"pending a validation pass".

Checklist

• I have installed pre-commit and installed the hooks with pre-commit install before creating any commits.
• I have updated any related tests for code I modified or added new tests where appropriate.
• I have manually tested my changes as follows:
  • Rewrote MPCAutofill/cardpicker/tests/test_local_art_edge.py for the new
    layout_class-driven comparison: border-colour-match → framed (black
    and white), the defect reproduction (dark off-hue artwork does NOT read
    extended), a genuine colour-mismatch positive (extended), the
    borderless short-circuit (proved via a monkeypatch that raises if
    pixel sampling is ever reached), an ambiguous-layout_class abstention,
    a two-directional mutation-proof test on the distance threshold, the
    pre-existing coordinate-frame asymmetry tests, and degenerate-input
    abstention. test_local_art_edge.py alone: 23 passed. Full
    cardpicker/tests/ suite (order-dependence, issue Test suite is order-dependent: leaked fetch-failure window trips the envelope across files (8 failures on master) #679): 3678 passed, 8
    skipped, 0 failed.
  • Built a one-off, non-committed synthetic-image probe
    (/tmp/opencode/scratch_art_edge/probe3.py) reproducing the pre-retune
    bug by hand against the code this PR replaces, confirming the retuned
    code no longer exhibits it.
  • Built a one-off, non-committed Scryfall validation probe
    (/tmp/opencode/scratch_art_edge/scryfall_validation.py) fetching 70
    real Scryfall card images across three ground-truth cohorts (rate-limited
    via the existing SCRYFALL_REST/SCRYFALL_CDN destinations, no image
    pixels persisted to disk or DB at any point) and running both the
    pre-retune and retuned classifiers against identical inputs — see the
    report for the full numbers and root-cause trace.
  • docs_lint.py, check_extractor_manifest_sync.py,
    check_extractor_ownership_totality.py all clean (unaffected, since no
    persistence/extraction code changed).
• I have updated any relevant documentation or created new documentation where appropriate.
  • docs/identification-pipeline.md and docs/pipeline-fidelity-gate.md's
    calculator roster entry updated to record the measured (negative) gate
    result instead of "pending a validation pass".
  • docs/reports/2026-08-06-art-edge-relative-comparison.md — the full
    measurement writeup.