Add Stage D fallback channel calculator (pre-fire prep)

[WilfordGrimley]

Description

Pre-fire prep for the full-catalog Stage D fire (code only — this PR does not run the fire, the targeted re-extraction of #340's 373-card cohort, or any other prod extraction/write; both remain separate owner-gated prod steps, see docs/pipeline-fidelity-gate.md).

Fallback channel calculator (calculate_fallback_verdict/run_fallback_calculator, own anonymous_id="stage-d-fallback-v1"): Stage D's own port of local_fallback.py's pilot "Pass 2" border/artist/symbol evidence-combination model, run only over cards the join-key calculator already found no confident hit for. Unlike the pilot's own live-image version, this operates entirely off already-persisted ImageEvidence fields:

• border: evidence.layout_class → local_fallback.filter_by_border_color (PROTECTED CORE, called not modified)
• artist: evidence.artist_ocr_name → local_fallback.match_artist (PROTECTED CORE, called not modified)
• symbol: evidence.symbol_phash → new _filter_by_symbol_phash, which calls local_fallback.render_set_symbol (PROTECTED CORE) and reimplements only the Hamming-distance arithmetic — the same pattern _symbol_phash_tiebreak already established for the join-key calculator, duplicated rather than shared since the two return different shapes (see that function's own docstring).

A vote is cast only when the intersection across every sub-check that produced a reading narrows to exactly one candidate — local_fallback.py's own documented rule, reproduced exactly, with no added agreement/corroboration layer (that layer doesn't exist in local_fallback.py, so this is a faithful port, not an augmented one). source=VoteSource.OCR (not DEDUCTION) — VoteSource's own docstring explicitly names "the border/artist/symbol evidence-combination fallback" as part of OCR's umbrella, not deductive_backfill's zero-image-inspection deduction.

Wired into the management command between the join-key and slow-path calculators (own verify_zero_resolutions gate check, same as join-key's), and _slow_path_eligible_cards_queryset now also excludes any card this calculator successfully voted on, so a card it resolves isn't also routed to human review in the same invocation.

Revision (2026-07-22): an earlier revision of this PR also added a .exclude(printing_tags__source=VoteSource.DEDUCTION) clause to the shared eligibility queryset ("constant #3" from the pipeline-fidelity gate's knowledge-inventory sweep — the pilot never re-voted a card the 2026-07-14 deductive backfill had already voted for). Owner ruled this OUT after a read-only investigation: that backfill is pure name/metadata deduction (never phash/OCR), its votes are sound (15-card sample all correct), and excluding those cards would strand ~27,819 sound-but-UNRESOLVED cards outside Stage D for no protective benefit — re-evaluating them is safe under the human-backed consensus gate (agreement dedups, disagreement surfaces to human review). The pilot's own exclusion was a performance optimization, not a soundness mechanism. That clause and its tests have been removed; the pre-existing per-calculator stable-anonymous_id idempotence exclusion (independent mechanism) is untouched. This PR is now fallback-calculator-only.

Constants #1 (RESOLUTION_FLOOR_DPI) and #2 (EXCLUDED_RESOLVED_TAGS) remain out of scope — their forward-impact sizing is still open per the gate page.

Protected core imported/called, never touched: local_fallback.filter_by_border_color, match_artist, render_set_symbol, SYMBOL_DISTANCE_THRESHOLD, SYMBOL_MARGIN, FALLBACK_CONFIDENCE_MULTI_EVIDENCE, FALLBACK_CONFIDENCE_SINGLE_EVIDENCE. local_fallback.py itself has zero diff in this PR.

Docs updated in the same PR: docs/features/catalog-completion-plan.md (Stage D "Pre-fire prep" entry, revised to record the constant-#3 non-restoration), docs/pipeline-fidelity-gate.md §3 (item 3 reframed as "intentionally not restored" with the owner-ruling rationale), docs/reports/2026-07-22-knowledge-inventory.md (constant-#3 row updated to match). No wiki update needed (backend-only, no user/admin-facing surface change) and no extractable-primitives.md row (hard dependency on resolve_and_persist_printing/CardPrintingTag, not a clean primitive).

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:
  • Ran the full backend suite from the pre-provisioned host venv (/home/ubuntu/.venvs/mpcautofill-pilot) against the live Postgres/ES containers' exposed ports (pytest-django's own ephemeral test DB, no prod read/write): 1642 passed, 4 skipped (the same CI-documented named skips), 0 failed.
  • mypy --config-file mypy.ini clean on both changed source files.
  • pre-commit run (ruff/isort/black/mypy/prettier) clean on every changed file.
  • makemigrations not applicable — zero model changes (confirmed models.py has no diff in this PR).
  • docs_lint.py clean.
  • Idempotence re-verified WITHOUT constant Simplify flag icons: static SVG assets instead of hand-rolled geometry #3: TestRunJoinKeyCalculator/TestRunSlowPathCalculator/TestRunFallbackCalculator's own test_idempotent_against_its_own_anonymous_id all still pass — idempotence comes entirely from the stable per-calculator anonymous_id exclusion, independent of constant Simplify flag icons: static SVG assets instead of hand-rolled geometry #3.
• I have updated any relevant documentation or created new documentation where appropriate.