Wire the attribute-chip casters into an engine; strip the vote from the uncalled wrapper

[WilfordGrimley]

Item 2 of the "essential major blockers" brief, sourced from PR #616 / docs/reports/2026-07-29-pipeline-coverage-composition-audit.md §1 Q1 and §6.

The defect

cast_border_attribute_vote, cast_frame_style_vote, cast_bleed_edge_vote in local_fallback.py are called from exactly two places:

1. local_identify_printing_tags.run_pilot — the live-FETCH pilot, with one completed run in its entire history (2026-07-16).
2. image_evidence.extract_card_evidence — which has ZERO production callers. Both engines call compute_card_evidence + persist_evidence directly and have since the 2026-07-20 fetch/compute decoupling.

So no engine could cast any attribute chip. Measured consequence: frame-style and bleed-edge chips at zero rows with no substitute. Border chips survived only because layout-class-cast-v1 independently computes the same thing — and that caster was reachable only from its own standalone management command, so the conveyor could not produce a border chip either.

The fix

New cardpicker/local_attribute_chip_cast.py — an evidence-reading caster, modelled on local_layout_class_cast (which reads evidence) and not on the pilot (which fetches). Zero image fetches, booby-trapped in test. Reuses classify_frame_style / FRAME_STYLE_TO_TAG / FRAME_VOTE_CONFIDENCE / BLEED_EDGE_TAG_NAME / BLEED_EDGE_VOTE_CONFIDENCE verbatim so it cannot drift from the pilot on what a frame class is or what it is worth.

Derivable populations (read-only, 2026-07-29): Modern Border 133,627 + Old Border 9,006 + appropriate-bleed 2,786 = 145,419 votes, none needing a fetch.

It does not cast border. That would be a third border channel, and the duplication is the audit's own §6 cull recommendation, not a pattern to extend. local_layout_class_cast remains the border caster; this PR makes it conveyor-reachable instead of duplicating it.

Two identities (frame-style-cast-v1, bleed-edge-cast-v1) written from one pass. The bleed chip is negative-only (~98.7% of cards get a frame vote and no bleed vote), so under one shared identity that frame vote would read as "handled" and permanently strand the bleed chip.

REQUIRED_EXTRACTOR_KEYS per chip family. The frame chip gates on artist_ocr as well as collector_line_ocr, because illus_anchor_fired is nullable and bool(None) is False — indistinguishable from "the extractor ran and found no anchor". Without the gate every card missing artist_ocr reads modern. That is item 4's bug class, applied correctly from the start.

Wired into the conveyor

stage_e_dispatch._run_stage_d now runs both casters after the four printing calculators. Zero fetches means no envelope or fetch-budget cost. Order is irrelevant here (neither reads another calculator's output) and the docstring says so.

A missing tag seed is caught and logged at ERROR rather than propagating: by that point the printing votes are already written, and letting a RuntimeError out would mark the whole dispatch FAILED over an advisory chip. Not a silent swallow — the counters staying at zero is the audit's own "this channel never ran" signal, and the log names the fix.

extract_card_evidence resolved

Renamed fetch_and_compute_card_evidence_for_tests and its cast_border_attribute_vote(...).save() deleted, per the brief's first option. The docstring now carries a DO NOT ADD A WRITE HERE note explaining why.

Its test class inverted with it. test_classified_border_casts_one_vote_per_card was a green test over a channel that had not existed for four months — a vote cast in an uncalled function looks like a wired channel to any grep, and a passing test looks like proof it works. That is how this stayed invisible.

Verification

mutation	expected red	result
drop the artist_ocr gate (read bool(None) as a real False)	frame chip skips rather than reading False	1 failed, 18 passed
unwire the casters from _run_stage_d	both conveyor-wiring tests	2 failed, 17 passed
restore the deleted vote cast	test-only-wrapper test + the inverted test_image_evidence class	2 failed, 129 passed
collapse to one shared anonymous_id	write / cross-suppression / idempotence	3 failed, 16 passed
let the seed RuntimeError escape	unseeded-tag dispatch test	1 failed, 19 passed
restored	—	20 passed

The third mutation initially came back green — the test stubbed the fetch to None, so layout_class was "" and the deleted cast would have returned None anyway. The test was rewritten to force a confident black reading, and only then did the mutation go red. Noted because a vacuous test is the same class of problem this whole PR is about.

• pytest cardpicker/tests/ — 3444 passed, 11 skipped (full suite)
• docs_lint.py --strict — clean
• pre-commit (ruff/isort/black/mypy/prettier) — clean
• No migration.

Two existing tethers fired and were satisfied, not bypassed: the calculator roster tether (docs/pipeline-fidelity-gate.md) and test_skip_reason_roster.py. Both new identities and all four new skip-reason constants are registered.

Docs (living, per .github/wiki-publish-map.json — no dated report)

• docs/features/printing-tags.md — new "Who actually casts the attribute chips" section with the caster/identity table.
• docs/pipeline-fidelity-gate.md — both new identities in the calculator roster, marked 0 rows / never run.
• docs/reference/skip-reasons.md — new section for the caster's four values.

🤖 Generated with Claude Code

https://claude.ai/code/session_013NhYmT1PxCcyemA16dFDxN