#47: close out the cutover; D5 restated: Processor owns the pin; goldens from co-core directly

[gregoryfoster]

Closes out the cutover and restates D5 now that watcher#350 has deleted Watcher's extraction. These are the operator decisions of 2026-10-10, from #47's hand-off.

Trap 1: the new generator equals the old oracle

scripts/gen_parity_goldens.py is rewritten in place, so its path is unchanged and tests/test_drift.py needs no edit. It drives co-core's pure extract API directly and follows the path Watcher took at a5d6f34:

• essence → extractor_for_essence;
• config is {**extraction_config_from_spec, **extraction_overrides_for_essence};
• the text is canonical_text, and its fingerprint is taken even when the text is empty;
• a spec_fingerprint that raises becomes None.

On co-core 0.19.7 it reproduces all 13 committed goldens byte for byte. test_reproduces_every_committed_golden was green before goldens.json changed at all, and it stays as a standing test.

Tests and fixtures

• Cross-implementation (trap 2): a test walks the script's AST. The script imports only the stdlib and co_core.pure.extract.*, nothing from processor.
• Metadata (trap 3): the top level of goldens.json is now co_core, generated_by and goldens. Watcher's generated_by, watcher_commit a5d6f34…, watcher_processor_version 0.19.7+1 and co_core moved into v1_provenance. That move was a one-time hand edit, and no digest was touched. Every regeneration carries v1_provenance forward and refuses to run without it.
• The bump note: each run prints every moved digest (case, field, old → new) and every added or removed case.
• test_extract.py: the version check now asserts PROCESSOR_VERSION == f"{GOLDENS['co_core']}+{LOCAL_GENERATION}". The docstring now names the new golden source.

TDD evidence (CR 6). The red step wasn't committed separately; b1e6369 carries the test and the generator together. The reds seen in-session: the new test module failed to collect (ModuleNotFoundError: No module named 'src', the old generator), and after the metadata moved, test_processor_version_matches_watchers raised KeyError: 'watcher_processor_version'. CR 4's red is in its own run: a bare ValueError that didn't name the case.

Docs

• Spec:
  • an amendment entry for 2026-10-10;
  • D5's row restated, with the lockstep history kept;
  • §5's bump rule: Processor decides, gives Watcher notice whenever processor_version moves, and coordinates only on a contract change;
  • §6 steps 4 and 5 marked done, linking watcher#326's gate verdict, its soak verdict, and watcher#350;
  • §7's golden source.
• AGENTS.md lines 21 and 55. The rule "never regenerate to make a test pass" stays.
• DEPLOYMENT.md: the co-core bump procedure.
• Plan: docs/plans/2026-10-10-47-post-cutover-d5.md.

Not changed

• src/. No deploy is needed: the drift check sees nothing that runs.
• The real corpus (real/, Real-input parity corpus from Watcher's export; record watcher#325's answers in the spec #16).

Verification

• uv run pytest: 597 passed, 6 skipped (after the CR round). The skips are the usual unreadable-credential denial cases.
• ruff check and ruff format --check are clean.

Remaining for #47 (off-repo)

• Take co-core's XML and feed extractors in step with Watcher, and re-run parity with a feed fixture #29 and JSON: take co-core's JsonExtractor (cannobserv#354) and add a jsonpath parity fixture #37 re-entry tests restated.
• The Watcher notice of decision 1 (drafted; posts only on approval).
• Become the cohort text-extraction processor: consume content.process, extract from one source_spec, store canonical text by hash, publish content.derived #1 closed after merge.

Refs #47, #1, #29, #37, epic #24, CannObserv/watcher#326, CannObserv/watcher#350.

🤖 Generated with Claude Code