Persist scryfall_cache volume and fail loud on a missing cache

[WilfordGrimley]

Description

Closes #402. The Scryfall bulk-data cache (MPCAutofill/scryfall_cache/default_cards.json, ~558MB) lived as a plain file inside the django/worker container filesystem with no persistent volume — every image rebuild silently destroyed it. Deploy-2 (2026-07-23) did exactly that, and local_calculate_verdicts went on to run to COMPLETED with a silently degraded (empty) back-face lookup, no error anywhere.

Two independent fixes, both requested by the issue:

1. Persistent volume: docker/docker-compose.prod.yml gains a named scryfall_cache volume (mirroring the existing postgres_data/elasticsearch_data pattern), mounted on both django (manual exec-run commands) and worker (the weekly import_canonical_card_data/update_dfcs qcluster schedule also reads/refreshes this same file — confirmed via mtg.py's get_canonical_cards_and_artists). The dev compose (docker/docker-compose.yml) only defines the postgres/elasticsearch services, not django/worker, so it doesn't have this hazard and needed no change.
2. Fail-loud guard: printing_metadata_import.ensure_scryfall_cache_present() is a new existence-only check (deliberately independent of the existing weekly _is_stale staleness/re-download logic, which is untouched), called at the very start of local_calculate_verdicts's Command.handle() before any card-by-card work. Raises a CommandError naming the missing path unless --allow-missing-scryfall-cache is passed explicitly — a message distinct from get_back_face_names's existing soft "empty set" per-card path, which stays as-is (still the right behaviour for a single lookup).

Deploy-time activation note: the compose change only takes effect at the next up --build -d deploy — it does not retroactively attach to the currently-running mpcautofill_django/mpcautofill_worker containers. No containers were rebuilt or restarted from this session.

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:
  • docker compose -f docker-compose.prod.yml config --quiet validates the compose file cleanly (no volume/syntax errors).
  • New tests: TestScryfallCacheGuard (5 tests) covering the guard directly (missing/present file) and wired through the management command (refuses without the flag, proceeds with the flag, proceeds when the cache is genuinely present). The 4 pre-existing TestCommandLedgerHardeningAndDryRunGuard tests were updated to pass --allow-missing-scryfall-cache since the test environment has no real cache file on disk (they're testing the dry-run guard, not this one).
  • Full backend suite: 1904 passed, 4 skipped (the 4 skips are the pre-existing documented named skips) via pytest cardpicker in /home/ubuntu/.venvs/mpcautofill-pilot.
  • pre-commit run (ruff/isort/black/mypy/prettier) clean on the final commit; docs_lint.py clean.
• I have updated any relevant documentation or created new documentation where appropriate.
  • docs/troubleshooting.md: new symptom-first entry for this incident (silently degraded back-face lookup after a rebuild), covering both the cause and the two-part fix, so a future occurrence (e.g. a box rebuild that skips the compose file) is diagnosable quickly.

Notes for reviewer

• No live containers were touched — this is code + compose-file only. The persistent volume starts empty on its first mount; the next import_scryfall_printing_metadata/import_canonical_card_data run (already scheduled weekly via worker's qcluster) will populate it, same as any fresh bootstrap.
• No migration involved; checked docs/troubleshooting.md/docs/infrastructure.md for an active migration freeze per CLAUDE.md's process — none found, and N/A here regardless (no schema change).