RFC: census-watch workflow with issues: write

[bong-water-water-bong]

Problem

The daily new-model watcher (tools/census_watch.py, merged with #242) alerts when a newly published Hugging Face model uses an architecture class that registry/architectures.json does not map. Its only alert path today is a systemd timer on the development box, which marks its unit failed. Nobody sees that unless they look at the box. The old watcher was dead for a week (203/EXEC) and nothing turned red.

Design

Add .github/workflows/census-watch.yml:

• Trigger: 04:30 UTC daily, plus workflow_dispatch. No pull_request trigger.
• Runner: ubuntu-latest, stdlib Python only, actions pinned by SHA.
• Action: runs scripts/census-watch.sh --limit 120.
• Alert: when the watch step fails, it files one deduped issue labelled census-watch. It does nothing while such an issue is still open.
• Permissions: contents: read, issues: write. The issues: write scope is the reason for this RFC.

The full file is in #242's first commit (b5a9d01). The follow-up PR would restore it unchanged.

Alternatives

• Timer only (today): silent unless someone looks at the box.
• Red check only (issues: read): a failed scheduled run notifies only whoever last edited the workflow.
• Post to the Discord #traffic-report bot instead of an issue: keeps GitHub permissions read-only, but moves the alert outside the repo.

Testing

• ctest census_watch (offline) covers the watcher's contract.
• Before merging, a workflow_dispatch run on the branch checks the issue path. Mapping a fake class, or running with an empty registry, forces the alert.

NOTICE

Nothing new: actions/checkout and actions/setup-python are already used by ci.yml.

[bong-water-water-bong]

Decision: accepted, with tighter permissions than the sketch. Implemented in #263.

• Token scope. The workflow token defaults to permissions: {}. Only a separate alert job holds issues: write, and that job runs no repository code: it only calls gh on the watch job's output, and only when the watcher reported an uncovered class.
• Untrusted text. Hugging Face model ids and class names land in the issue inside a code fence, with backticks and @ stripped and at most 40 lines.
• Credentials and pins. Checkout doesn't persist credentials. The actions are the same SHA pins ci.yml already uses.
• Testing. A workflow_dispatch input simulate_alert exercises the issue path once after merge; the test issue gets closed.

Why an issue rather than the alternatives:

• an issue is the only option that's both visible in the repo and deduped;
• a red scheduled check only notifies whoever last edited the workflow;
• Discord would move the alert out of the repo.

The development box's systemd timer stays as a second path.

[bong-water-water-bong]

Landed. The workflow is in #263, and #265 tunes the alert rule.

• Alert path tested. A workflow_dispatch run with simulate_alert filed census: new HF model class uncovered by the registry #264 with the sanitized body.
• That run also found real noise. It listed three custom-code research classes (auto_map, one uploader each). census-watch: custom-code one-offs are reported, not an alert #265 reports classes like that without alerting until 3 uploaders use one. Native and reviewed classes alert as before.
• Plain run green. A plain dispatch after both merges: the watch job succeeded, the alert job was skipped, and no issue is open.

The first scheduled run is tomorrow at 04:30 UTC.