feat(stage-d): wire four FREE evidence-only vote channels into the conveyor

[WilfordGrimley]

Summary

Wires four FREE-data vote-casting calculators into the streaming conveyor's
stage_e_dispatch._run_stage_d (and therefore into run_pipeline's bulk
monolith, which calls the same function): ai-art-detector-v1,
lands-artist-decomp-v1, residual-classify-v1, art-hash-artist-v1. Each
had a real, tested, card_ids-scoped code path already (issue #533's own
batch-scoping prerequisite already applied to each) but no caller reachable
from the conveyor — a channel that never runs and a channel that runs and
casts zero votes are indistinguishable from the outside, which is the
defect this closes.

Roster count correction: the derived calculator roster
(docs_lint.py::check_calculator_roster_tether) is 17 real vote-casting
identities (19 declared, 2 allowlisted as non-calculators), not the ~28
previously circulated. 11 of 17 are now invoked by a real call site (7
before this PR, 4 added here); snip 6 are deliberately left unwired with a
stated reason and a tracked issue.

Roster table (the answer to "does a pass invoke this")

Identity	Wired before	Wired now	FREE/EXPENSIVE	Measured cost
stage-d-join-key-v1	yes	yes	FREE	stored evidence only
stage-d-fallback-v1	yes	yes	FREE	stored evidence only
stage-d-illustration-v2	yes	yes	FREE	stored evidence only
stage-d-slow-path-v1	yes (router, 0 votes)	yes	FREE	CardScanLog only
layout-class-cast-v1	yes	yes	FREE	stored evidence only
frame-style-cast-v1	yes	yes	FREE	stored evidence only
bleed-edge-cast-v1	yes	yes	FREE	stored evidence only
ai-art-detector-v1	no	yes	FREE	3 stored OCR text fields, regex scan
lands-artist-decomp-v1	no	yes	FREE (fetch_budget=0)	evidence-backed only; snip index cached 1.48s once/process
residual-classify-v1	no	yes	FREE (both refetch budgets=0)	phash-only path; snip index cached, lazy
art-hash-artist-v1	no	yes	FREE	pure DB read (content_phash + resolved artist chain)
art-edge-continuity-v1	no	no	FREE data, but gated	own stated validation precondition unmet — #721
deductive-backfill-v1	no	no	EXPENSIVE (implementation)	no card_ids param; snip rebuilds 113k-row index/call — #722
local-name-frequency-v1	no	no	EXPENSIVE (implementation)	no card_ids param; snip catalogue-wide census/call — #722
local-ocr-v1	no	no	EXPENSIVE (data source)	primary caster (run_pilot) needs live fetch+tesseract — #723
local-phash-v1	no	no	EXPENSIVE (data source)	primary caster (run_pilot) needs live fetch+hash — #723
local-fallback-v1	no	no	EXPENSIVE (data source)	primary caster needs live fetch — #723

(evidence-transfer-v1/question-feed-hypothetical-vote are allowlisted —
not vote-casting calculators, see CALCULATOR_ROSTER_ALLOWLIST.)

What shipped

1. stage_e_dispatch.py: new _run_evidence_only_calculators(run_id, card_ids, outcome, dry_run), called from _run_stage_d after the
   attribute-chip casters, with a new seam("stage-d:evidence-only")
   envelope-check point matching the existing convention. Calls
   run_ai_art_detector, run_frame_mismatch_recovery
   (ocr_refetch_budget=0, fallback_refetch_budget=0),
   run_d0_sibling_artist_propagation, run_lands_identify
   (sample_size=None, fetch_budget=0) — every live-fetch budget forced to
   0, exactly the "scoped, genuinely free" path each module's own docstring
   already documents. run_ai_art_detector's RuntimeError (missing
   "AI-Generated" Tag seed) is caught and logged, matching
   _run_attribute_chip_casters' own established pattern — a missing
   operator seed must not fail votes the printing calculators already
   wrote this batch. Six new DispatchOutcome counters.
2. stage_e_dispatch.dispatch_micro_batch: the six new counters folded
   into the streaming dispatch's PilotRunLedger.counters. Also added the
   three PRE-EXISTING stage_d_border_chip_votes/frame_chip_votes/
   bleed_chip_votes counters to this same merge — they were present on
   DispatchOutcome and tested against it directly, but were never
   surfaced into the streaming ledger's own counters dict (bulk mode's
   run_pipeline.py already reported them). Minor, directly adjacent fix,
   folded into the same merge dict this PR was already touching.
3. run_pipeline.py: the 6 new + (belatedly) 3 chip counters added to
   both the bulk-mode result dict and the streaming-mode acc
   accumulator, plus the STAGE D log line text updated to name the four
   new calculators.
4. test_stage_e_dispatch.py: new TestEvidenceOnlyCalculators, one
   test per channel, each asserting on dispatch_micro_batch's own
   DispatchOutcome/PilotRunLedger — never on the calculator function in
   isolation (that coverage already exists in each calculator's own test
   module). A fifth test proves run_ai_art_detector's missing-tag-seed
   RuntimeError does not halt the other three channels in the same
   batch.
5. docs/pipeline-fidelity-gate.md: new "Wiring status" section (the
   roster table above, in durable form) plus issue links added to the
   existing art-edge-continuity-v1/local-name-frequency-v1 roster
   entries.
6. Three new issues with concrete acceptance criteria for the 6
   deferred channels: #721 (art-edge-continuity-v1),
   #722 (deductive-backfill-v1 + local-name-frequency-v1, shared root cause),
   #723 (local-ocr-v1/local-phash-v1/local-fallback-v1, shared root cause).

Deviations from spec

• Consolidated 6 deferred channels into 3 issues, not 6. Each pair/trio
  shares one root cause (a missing card_ids scoping prerequisite for
  deductive-backfill/name-frequency; snip the dormant live-fetch pilot engine
  for ocr/phash/fallback) and one fix would resolve both/all three members
  at once — a separate issue per identity would fragment one piece of work
  across multiple trackers. Each issue still states per-channel acceptance
  criteria.
• Did not modify local_art_edge.py/deductive_backfill.py/
  local_identify_printing_tags.py to point at the new issues from inside
  their own docstrings. The brief asks for a reason "in a durable place
  (the module's own docstring, OR a doc under docs/)" — docs/ pipeline-fidelity-gate.md's new section satisfies this without touching
  three additional source files for a documentation-only cross-reference.
• The "10 of roughly 28 channels" premise in the brief was wrong on both
  numbers, not just the wired count — the real roster is 17, not ~28.
  Reported as found, not corrected silently.
• The brief's "established facts" section (three calculators, no
  run_slow_path_calculator/no attribute-chip casters) was already stale
  relative to current master — PR Wire the attribute-chip casters into an engine; strip the vote from the uncalled wrapper #654 (2026-07-30, already on master
  before this branch was cut) had already wired run_slow_path_calculator
  and the three attribute-chip casters. This PR's roster table reports the
  true current state rather than the brief's premise.

Test plan

• python -m pytest cardpicker/tests/test_stage_e_dispatch.py — 93
  passed (individually)
• python -m pytest cardpicker/tests/test_run_pipeline.py — 32
  passed (individually)
• python -m pytest cardpicker/tests/test_stage_e_dispatch.py cardpicker/tests/test_run_pipeline.py — 125 passed (together)
• python -m pytest cardpicker/tests/test_local_detect_ai_art.py cardpicker/tests/test_local_lands_identify.py cardpicker/tests/test_local_residual_classify.py — 175 passed
  (individually and together; snip these calculator modules were not
  modified, verified as a due-diligence check)
• All five files together — 300 passed, matching 93+32+175 exactly
• Mutation check: each of the 5 new dispatch-level tests verified
  to go red when _run_evidence_only_calculators(...) is removed
  from _run_stage_d, then the call restored and the same 5 verified
  green again
• pre-commit run (ruff, isort, black, mypy, prettier) — clean on all
  4 changed files
• python .github/scripts/docs_lint.py — clean (roster tether intact)
• Pre-existing, unrelated: test_artbox_exemplar_backfill.py errors at
  SETUP with AttributeError: type object 'PytestDjangoTestCase' has no attribute '_pre_setup_ran_eagerly' — a documented pre-existing
  environment issue, not touched by this PR, not run as part of the
  verification above

Backfill (NOT run as part of this PR)

Wiring affects future passes only. The existing catalogue needs a
catch-up run per newly-wired channel, over stored evidence:

• ai-art-detector-v1: management/commands/local_detect_ai_art.py
  already exists and calls run_ai_art_detector in exactly the same
  card_ids=None / stored-evidence-only shape this PR wires per-batch — it
  suffices as the backfill command, no new command needed.
• lands-artist-decomp-v1: management/commands/local_lands_identify.py
  already exists and calls run_lands_identify, but its CLI defaults
  (sample_size=300, non-zero fetch_budget) target the pilot's own
  sampled/fetch-budgeted mode, not the free evidence-backed path this PR
  wires. A catch-up run needs --sample-size=<None-equivalent> --fetch-budget=0 (or equivalent full-pool, zero-fetch flags) to match
  what the conveyor now does per batch — check the command's current flag
  surface before running.
• residual-classify-v1/art-hash-artist-v1:
  management/commands/local_residual_classify.py already exists and
  calls both run_frame_mismatch_recovery/run_d0_sibling_artist_propagation.
  For the phash-only free path this PR wires, run with
  --ocr-refetch-budget=0 --fallback-refetch-budget=0 (or equivalent) to
  match; snip a full run with non-zero refetch budgets would ALSO cover the
  OCR/fallback-flagged rows this PR intentionally leaves for the live-fetch
  pilot decision tracked in local-ocr-v1 / local-phash-v1 / local-fallback-v1: the live-fetch pilot engine channels remain unreachable from the conveyor #723, and is a separate, larger-scope backfill
  decision, not what this PR's wiring covers.

Sequencing caveat carried over from the brief: ai-art-detector-v1
and the evidence-backed branches of lands-artist-decomp-v1/
residual-classify-v1 all read Stage C OCR text. A full-catalogue pass
was reported live on this box re-deriving that same OCR text under new
extractor versions at the time this branch was cut — a backfill run for
any of these three channels started before that pass completes would
produce verdicts against stale extractor output that would need
re-running once it finishes. Do not run any backfill from this PR
without first confirming that pass has completed.

Open items

1. Should art-edge-continuity-v1's validation pass (art-edge-continuity-v1: run the stated extended-art validation pass before wiring the vote #721) run before or
   after this PR merges? It needs no live fetch (pure DB read against
   already-imported Scryfall frame_effects), so it could run independent
   of the concurrent full-catalogue pass mentioned in the brief — owner
   call.
2. deductive-backfill-v1/local-name-frequency-v1 (deductive-backfill-v1 / local-name-frequency-v1: add card_ids batch scoping so these FREE-data channels can be wired into Stage D #722): is adding
   card_ids scoping + the shared process-cache to deductive-backfill-v1
   worth doing, or should this channel be retired instead (issue deductive backfill: the first tier's "cross-verified against Scryfall" condition excludes nothing (137 -> 137) #600
   already flags that its D1 tier's advertised external corroboration was
   never implemented)? local-name-frequency-v1's own roster entry already
   says "may be retired" — this PR does not resolve that question, only
   sharpens why it's currently unwireable.
3. local-ocr-v1/local-phash-v1/local-fallback-v1 (local-ocr-v1 / local-phash-v1 / local-fallback-v1: the live-fetch pilot engine channels remain unreachable from the conveyor #723): is the
   dormant run_pilot engine still a channel this project wants live at
   all, given local_calculate_verdicts' three calculators already cover
   the equivalent ground via Stage C's evidence-first design? This is
   the largest open question of the three issues and needs an owner
   decision before any further work on it.

Live state

• Branch wire-free-vote-channels pushed to origin, 1 commit ahead of
  master (17345a48).
• No PR merge attempted — pushed and reporting only, per instructions.
• Issues art-edge-continuity-v1: run the stated extended-art validation pass before wiring the vote #721, deductive-backfill-v1 / local-name-frequency-v1: add card_ids batch scoping so these FREE-data channels can be wired into Stage D #722, local-ocr-v1 / local-phash-v1 / local-fallback-v1: the live-fetch pilot engine channels remain unreachable from the conveyor #723 opened on
  ProxyPrints/ProxyPrints.github.io, all open, no other action taken on
  them.
• No management command, backfill, or write against the production
  database was run at any point in this session.