[Deicyde]

Hold until release B is promoted. site/ops/brain-nightly.sh is release-affecting. Merging this while B is being built at the main you're preparing (after #102, #92 and #93) would force a rebuild. It goes into the next release cycle.

Problem

brain-nightly.sh refuses to publish while release-affecting source is uncommitted. Its pathspec list misses four modules that the reducer and the release tools import, so a local edit to one of them was frozen under a Git authority that lacked it:

• brain/install_wikidata_observation.py and brain/wikidata_observation.py, imported by brain/build_common.py
• catalog/huggingface_download.py, imported by brain/build_common.py
• brain/tools/execution_environment.py, imported by the release tools

site/ops/brain_release_authority.py already treats all four as release-affecting. Only the nightly's source check missed them. Found in #75's re-review.

Change

• The four paths join the uncommitted-source check.
• New test, test_release_cleanliness_gate_covers_every_module_the_reducer_imports (site/ops/test_brain_nightly.py):
  • It computes the local-import closure of the reducer scope (reducer-inputs-v1.json) and the release tools (build_release, verify_release, measure_store, authority_contracts, fold_proposals, the Wikidata bundle modules and ingest/common). Lazy imports inside functions count, and imports resolve along the scripts' sys.path (brain, brain/tools, site, catalog, site/ops).
  • It then requires every module in that closure to be covered by the nightly's pathspecs, so a future import can't slip past.
  • Today the closure has 24 modules; on main the test would fail on exactly these four.

bash -n passes under bash and /bin/bash 3.2. I didn't run the suite locally because machine load was above the agreed limit; CI runs it.

[Deicyde]

Verdict: merge after release B is promoted, as the description says. Nothing blocks merge.

• The import closure: I re-ran the new test's import walk against both checkouts.
  • At this head the reducer's closure has 24 modules, all watched.
  • At main 3c1139f exactly four are unwatched, the four the description names: brain/install_wikidata_observation.py, brain/wikidata_observation.py, brain/tools/execution_environment.py and catalog/huggingface_download.py.
  • The walk resolves the one relative import (from . import execution_environment in authority_contracts.py) through its level-0 fallback.
• The effect: a dirty file among those modules now sets DIRTY, and the existing check aborts the release. None of the four is release-neutral in brain_release_authority.py, so the claim that they already counted as release-affecting holds.
• Release impact: site/ops/brain-nightly.sh is in AFFECTING_EXCEPTIONS, so merging forces a rebuild of release B. It has to wait for promotion.
• CI: the python job ran site/ops/test_brain_nightly.py with 26 tests (25 on main), and the browser, worker and required jobs passed. The PR merges cleanly into main 3c1139f.

Posted by Claude Code (AutoformBot "Merge Currently Open PRs 4" session).