test(artifact-registry-coverage): Phase 4 PR 4 — ae-data producer-side CI guard

[cipher813]

Summary

Phase 4 of the artifact-freshness-monitor arc (plan doc at ~/Development/alpha-engine-docs/private/artifact-freshness-monitor-260527.md; ROADMAP P1 entry in alpha-engine-config #342 / #344). PR 4 of the 4-PR cascade across ae-data / ae-research / ae-predictor / ae-backtester. Mirrors the tests/test_schema_contract.py::_GRANDFATHERED_BARE_FIELDS precedent — force operator attention at every new producer addition so the absence-of-artifact bug class can't slip through new code without an explicit register-or-grandfather decision.

PRs 5-7 (ae-research / ae-predictor / ae-backtester) mirror this pattern in a follow-up session per the plan §8 estimate (~2 sessions for PRs 4-7).

Depends on:

• alpha-engine-config #344 (merged) — the registry SoT this test forces operator coverage against
• Independent of #335 freshness-monitor Lambda — this is the PR-time chokepoint; that Lambda is the runtime probe

Changes

tests/test_artifact_registry_coverage.py (new):

• _enumerate_put_sites() — git grep for every s3.put_object( / s3.upload_file( literal in ae-data production code (excludes tests/, infrastructure/lambdas/, .claude/, .venv/, build/).
• EXPECTED_PER_FILE_PUT_COUNTS — hand-curated pin of 44 PUT sites across 29 files as of 2026-05-27. Bumped on every legitimate PUT-site change.
• 3 tests:
  • test_every_producer_file_is_pinned — new file with a PUT site fails with a clear resolution path (register in ae-config registry → bump the pinned set).
  • test_every_pinned_file_still_exists — file removed without bumping the pin fails (stale-pin detection).
  • test_pinned_counts_match_actual — PUT-site count drift in an existing file fails (the typical add-a-new-PUT case).

Design choice — per-file count rather than per-key-template extraction

Statically extracting key templates from arbitrary f-string put_object(Key=...) calls is fragile (keys are often constructed from surrounding context: loop variables, function args, config values). Per-file count is stable across refactors and sufficient to force operator review at every new addition. The operator is the source of truth for "which key does this PUT emit"; the test's job is to ensure the operator can't sleepwalk past adding one without thinking.

This is the lighter-touch alternative to the plan's "per-key extraction + match against registered prefix" specification. Bias toward chokepoint robustness — the test should be noisy in the right way, never silent in the wrong way.

Sister coverage

• alpha-engine-config #344 → scripts/validate_artifact_registry.py — PR-time SoT schema validator (registry side).
• alpha-engine-data #335 → infrastructure/lambdas/freshness-monitor/index.py::load_registry — Lambda-runtime registry parse + spec validation (deploy side).
• This PR — producer-side coverage CI guard (PR-time).

Test plan

• pytest tests/test_artifact_registry_coverage.py -v — 3 passing
• Stress-test: add a synthetic s3.put_object(...) line to any existing producer file → test fails with the operator-resolution message
• Stress-test: add a synthetic new file with a put_object call → test_every_producer_file_is_pinned fails

Composes with

• [[feedback_lift_invariants_to_chokepoint_after_second_recurrence]] — this IS the chokepoint cascade for the absence-of-artifact bug class.
• [[feedback_no_silent_fails]] — operationalized at PR time for new producers.
• [[feedback_sota_institutional_default_no_shortcuts]] — operator-attention chokepoint over heuristic key extraction.

🤖 Generated with Claude Code