Retire PrintingTagVote: the vote channel that had no resolver, no reader and no rows

[WilfordGrimley]

Owner ruling, 2026-07-29: "i am willing to not need the printing tag. i am happy to reduce things to the minimum that gives us our expected results. we don't need to reinvent the wheel everyday." And on its origin: "printingtag was half cooked."

Implements PR #599's §7.1 / §8 recommendation. Every claim below was re-verified against production and the working tree by this PR's author, not taken on trust from the investigation.

---

Why this is safe

check	result
PrintingTagVote rows in production	0 (select count(*) from cardpicker_printingtagvote, read-only, 2026-07-29)
…of which human	0 — the table is empty of every source
consensus resolver	none exists. printing_tag_consensus.py, forward-referenced by two docstrings, is absent from disk, from git log --all, and from every open PR
readers	none outside the Django admin
frontend callers of submitPrintingTagVote	none anywhere in frontend/src/ (re-grepped)
PROTECTED CORE	not protected. docs/upstreaming/license-provenance.md §2 excludes models.py from file-level protection; its manual-review carve-out names exactly VoteSource, AbstractWeightedVote, CanonicalPrintingMetadata, CardPrintingTag. The three protected consensus modules do not reference PrintingTagVote. No license-provenance review is triggered.

Human votes are the one thing in this system that cannot be re-derived. There are none here, so there is nothing to migrate, export or rescue.

The deeper reason it was unsalvageable rather than merely unused: an imported Scryfall fact is not a disputable claim. resolve_weighted_consensus's has_human_backed gate is absolute and independent of the weight sum, so a machine-only channel returns None at any volume — the model was built to carry an indisputable fact through a mechanism designed for disputable ones.

What is NOT touched

The names are confusingly adjacent and an earlier orchestrator conflated two of them while briefing this change. Stated explicitly:

• CardPrintingTag — 167,206 rows live. "This catalogue IMAGE depicts this Scryfall PRINTING". The entire Stage D printing channel, read by printing_consensus into Card.printing_tag_status / inferred_canonical_card / Elasticsearch. Untouched.
• CardTagVote — 223,909 rows live. Read by tag_consensus into Card.tags. Untouched.
• PRINTING_TAG_MIN_VOTES / PRINTING_TAG_IMPLICIT_CAP / PRINTING_TAG_MACHINE_WEIGHT — app-wide consensus weights, never on a PrintingTagVote path. Untouched.
• _split_new_printing_tag_votes / _purge_and_write_printing_tag_votes — CardPrintingTag collision guards (list[CardPrintingTag] signatures). Untouched.

verify_no_machine_only_resolutions already excluded PrintingTagVote at purge_machine_votes.py:139 — confirmed; only its now-stale explanatory note changed, no gate logic.

What shipped

1. PrintingTagVote model deleted; migration 0101_delete_printingtagvote (DeleteModel).
2. views.post_submit_printing_tag_vote, _SubmitPrintingTagVoteRequest, _get_canonical_card_or_400, and the 2/submitPrintingTagVote/ route deleted. No frontend caller existed, so this is not a breaking API change in practice. (The rate-limit helpers named _printing_tag_rate_limit_* stay — five other endpoints share them.)
3. AdminPrintingTagVote deleted.
4. purge_machine_votes: the queryset, both .delete() calls, and PurgeResult.printing_tag_votes_deleted removed. The counter was dropped, not pinned at 0 — a purge report naming a table that no longer exists is a claim about work it did not do, and this module's whole premise is that its output must be sufficient on its own to reconstruct what was destroyed. The rationale is written into _report_counts' docstring beside the pre-existing "keep the wording stable for log scrapers" note it deviates from. test_purge_machine_votes.py's test_covers_all_four_vote_tables becomes test_covers_every_vote_table over three.
5. manage.py import_external_ip_tags (579 lines) and test_import_external_ip_tags.py deleted — see the ruling below.
6. test_submit_printing_tag_vote.py deleted.
7. Both phantom printing_tag_consensus.py forward-references are gone, along with the file that carried them.
8. Docstring corrections in vote_write.py and integrations/game/scryfall_bulk_data.py (whose "three importers" origin story now names two live ones and one retired).

The dormant importer — ruling: REMOVE, with the knowledge relocated, not the PR description

It could not have been kept. The command's entire purpose is writing PrintingTagVote rows; with the model gone it would not import, let alone run. "Keep it and make it fail loudly" would mean gutting its write path anyway, leaving a stub that preserves nothing.

So the knowledge was moved somewhere greppable from a checkout, not left in a PR description. docs/features/printing-tags.md's External-IP entry is rewritten from "here is how the import works" into a retirement record carrying:

• the reusable algorithm in full prose — the art_tags bulk entry, the BFS-to-fixpoint over child_ids from the external-ip slug (deliberately not the one level the original plan specified) with its fail-loud-on-missing-slug rule, the two-pass tag/tagging split that keeps the taggings payload out of memory during indexing, and the illustration_id -> default_cards (incl. per-face card_faces) -> CanonicalCard.identifier join;
• the measured numbers on both signals: promo_types reaches 10,407 of 113,224 printings at 100% per-set recall; art:external-ip closes over 2,799 tags / 8,332 illustrations reaching ~13,166 — an uncharacterised ~2,759 delta that is the whole case for the artwork-origin question being distinct;
• what a rebuild must not carry: the vote framing, the negative pass (~100k NOT_APPLICABLE rows into a table nothing read), the source=DEDUCTION/scryfall-tagger-v1/run_id/purge plumbing, and security_stamp as a UB signal;
• the grain finding (0 of 50,828 illustrations appear on both a UB and a non-UB printing, despite 25,381 reprinted and 20,542 crossing sets) — so an illustration-keyed attribute is the right shape;
• git show e6c6429a:MPCAutofill/cardpicker/management/commands/import_external_ip_tags.py to read the deleted code verbatim.

Two contracts were rescued rather than deleted with it:

• EXTERNAL_IP_TAG_NAME moved to reason_tags.py, beside the surviving human channel. It is a convergence contract — both channels must write one Tag.name so tag:external-ip is a single predicate — and it now outlives the code that used to honour it. test_reason_tags still pins it.
• vote_write's target_field parameter lost its only production caller. Its two tests are re-pointed at CardPrintingTag.printing_id (which also has that column) rather than deleted, so the parameter stays exercised instead of quietly becoming a card_id constant. The re-pointed test is sharper than the original: the stale row belongs to a different card, same printing, so only a printing_id-keyed purge can reach it.

scryfall_bulk_data.ART_TAGS is kept with its live-endpoint test although nothing now consumes it — one constant plus one test is cheaper to carry than re-deriving which bulk entry type the Tagger feed lives under, and the test is what would tell us if it disappeared upstream in the meantime.

Verification

• Production row count: 0, confirmed directly before writing the DeleteModel, alongside CardPrintingTag 167,206 and CardTagVote 223,909 as a positive check that the right table was identified.
• Full cardpicker/tests/ suite: 3237 passed, 8 skipped, 0 failed (4m16s, run twice).
• makemigrations --check --dry-run: "No changes detected", exit 0.
• Migration graph leaf count for cardpicker: exactly 1 ([('cardpicker', '0101_delete_printingtagvote')]), asserted via MigrationLoader.
• docs_lint.py --strict: clean.
• pre-commit on every changed file (ruff, isort, black, mypy, prettier, README parity): all pass.
• Grep sweep for PrintingTagVote, printing_tag_consensus, import_external_ip_tags, submitPrintingTagVote, scryfall-tagger: every surviving hit is a deliberate historical aside, plus the two TestSplitNewPrintingTagVotes / TestPurgeAndWritePrintingTagVotes class names that are about CardPrintingTag and must stay.

⚠️ Migration numbering — read before merging

The file is numbered 0101 per the orchestrator's allocation (#601 → 0099, #604 → 0100), but its dependencies point at 0098_card_illustration_consensus_fields, master's real and only leaf, so this branch is migratable and single-leaf today and its CI can actually run.

This follows the convention 0098_superseded_card_printing_tag_archive (#604) states in its own docstring: "a dependency on a migration that does not exist on master makes THIS branch unmigratable today, with certainty, in exchange for avoiding a collision that may never happen. A collision, by contrast, is loud and caught at merge time by makemigrations --check."

=> Whichever of #601 / #604 merges before this, repoint this migration's dependency at the new leaf on rebase. Leaving it on 0098 after 0099/0100 land gives cardpicker two leaves, and because pytest-django builds its test DB by running migrate, that takes every branch's CI down — the failure #576 exists to repair. The migration's own docstring says all of this at the top.

Current state confirmed 2026-07-29: #601 has already renumbered to 0099_rename_printings_count_catalogued; #604 is still at 0098 as 0098_superseded_card_printing_tag_archive and expected to become 0100.

⚠️ Sequencing against #588

#588 (open) adds a docs/pipeline-fidelity-gate.md roster entry for scryfall-tagger-v1 that links to management/commands/import_external_ip_tags.py — a file this PR deletes. If #588 merges first, that entry must be removed (or rewritten as retired) here on rebase, or docs_lint's path check will fail on the dead link. If this merges first, #588's recursive-glob fix will simply not find the identity and its roster row should not be added.

Per PR #599 §10.10 the recommended order is #588 first: the tether fix is correct independent of this ruling, and a row that is later deleted is cheaper than a tether that silently misses the next dormant identity. Either order works; they must not be merged blind to each other.

Deviations from the brief

1. Migration dependency on 0098 rather than the not-yet-existent 0100. See above. The brief also required verifying a single leaf and running the suite — impossible with a dangling dependency, since Django cannot build the graph at all. Both requirements are met this way; the merge-time repoint is documented in three places.
2. printing_tag_votes_deleted removed rather than pinned at 0. PR Decision doc: should PrintingTagVote exist? Printing-tag vs illustration-tag grain #599 §8.2 offered both. Reasoning is in _report_counts' docstring and above.
3. Two vote_write tests re-pointed rather than deleted, to keep target_field a tested parameter.
4. docs/reports/2026-07-29-fidelity-gate-recheck.md got a dated "PARTLY SUPERSEDED" banner rather than edits. It is a same-day audit whose PrintingTagVote/scryfall-tagger-v1 findings were accurate when measured and are part of what motivated this removal; rewriting a dated report to match a later decision is the opposite of what this repo's correction log exists for.
5. test_vote_consensus.py's "some other calculator" example swapped from scryfall-tagger-v1 to stage-d-illustration-v1. The assertion is about the three-conjunct zero-weight rule and is unchanged, but an example identity nobody can look up is a worse example.

Open items for the owner

1. Seed the external-ip Tag in production — manage.py seed_no_match_reason_tags. It does not exist live (Tag.objects.count() is 30), so the human card-grain channel — the one that works end to end — still cannot record external-ip. Independent of everything here, and one command.
2. Characterise the ~2,759-printing delta between art:external-ip and promo_types. A ~50-row sample decides whether the Tagger dependency survives into the unified importer at all. The algorithm to do it is preserved in the retirement record.
3. Surface UB from promo_types — 3,450 catalogue images already resolve to a UB printing today through the existing Card.canonical_card link, with no vote of any kind. Denormalising onto Card.tags gets Elasticsearch filtering free and shares one predicate with the human channel.

🤖 Generated with Claude Code

https://claude.ai/code/session_013NhYmT1PxCcyemA16dFDxN

[WilfordGrimley]

Rebased onto origin/master (4a3babc); dependency repointed to 0100. This PR's CI is red on purpose.

Merge order is fixed: #604 first, then this one. 0101_delete_printingtagvote now depends on 0100_superseded_card_printing_tag_archive (PR #604), which is not on master yet — so migrate on this branch alone raises NodeNotFoundError, every test errors at pytest-django's test-database setup, and #611's leaf guard reports two leaves. All five failing checks are that one cause. The red clears the moment #604 merges. Do not 'fix' it by repointing at 0099.

Why this order, and why one of the two branches had to be red

Exactly one of #604/#615 can own the dependency on master's real leaf (0099), so exactly one can be migratable at a time. The choice is which PR to leave unverifiable, not whether to leave one. #604 is the larger change and the Stage-D monolith's core, so it keeps the verifiable, green branch; this PR is small and cheap to hold.

The rejected alternative — depend on 0099, renumber this 0100 — makes this branch green immediately but leaves cardpicker with two leaves after both merge, which takes every branch's CI in the repo down at test-database setup. A local red costs one PR; a forked graph costs the repo. (The earlier docstring preferred the green-now convention on the then-true premise that nothing had claimed a number above master's leaf. #604 has now claimed 0100, which inverts that trade-off.)

Verified on a simulated merge, not by eye

master + #604 + this branch merges cleanly, and on that tree:

• MigrationLoader(None).graph.leaf_nodes() → exactly one cardpicker leaf, 0101_delete_printingtagvote. Chain: 0098 (Illustration consensus: md5 identity-group pooling and sibling propagation, built in from the first line #573) → 0099 (printings_count counts our own rows: rename it, and delete the "cross-verified against Scryfall" claim #601) → 0100 (Run-scoped eligibility, the value-comparing split, and an archive for superseded votes #604) → 0101 (this).
• Full pytest cardpicker/tests/: 3350 passed, 11 skipped.

Two real rebase breakages fixed (unrelated to migrations)

Master's #588 added a fidelity-gate roster entry and a real-repo tether assertion for scryfall-tagger-v1 — the one calculator identity declared under management/commands/, the subtree this PR deletes. Both went stale:

• docs/pipeline-fidelity-gate.md: roster entry rewritten as a retirement record (broken link to the deleted command removed, knowledge kept). Its adjacent "open owner call" on PrintingTagVote's absence from the gate is now marked resolved — answered by removal rather than by scoping.
• .github/scripts/tests/test_docs_lint.py: scryfall-tagger-v1 removed from the real-repo tether list, which no longer has a live management/commands/ witness. The recursion property it guarded stays covered on fixtures by test_scan_is_recursive_but_bounded.

docs_lint.py --strict clean; its 47 rule unit tests pass.

🤖 Generated with Claude Code