docs: the sweep is manual; Operations360 is where it's headed

[kurtseifried]

Follow-up to #49, which shipped a runbook claiming "a cloud routine runs the sweep weekly." It does not, and the doc needed to stop saying so.

What the test run found

I triggered the routine once #49 landed. It read the runbook, followed the spec, and died at step 2:

tool_use Bash: gh api repos/CloudSecurityAlliance-Internal/CSA-Plugins
tool_result ERROR: Exit code 127
  /bin/bash: line 1: gh: command not found

gh is absent from the cloud sandbox entirely — not a token-scope problem. Nothing at /usr/bin/gh, /usr/local/bin/gh, or /opt/homebrew/bin/gh. The sweep is built end-to-end on gh api, so it cannot run there as written.

Two things the run did prove, and they are why the structure below is worth keeping:

• runbook-as-spec works — the agent read docs/periodic-sweep.md and followed it without the prompt restating the steps
• the guard held — it stopped and reported inability rather than emitting a false all-clear

What changed

• Routine trig_01TQh4GMWKRnt4L4QpM5mhJc is disabled, not deleted, so the wiring survives if Operations360 uses the same mechanism.
• Running this as a scheduled routine → How this gets run + ### The job spec. The spec is now scheduler-agnostic: Operations360, a routine, or a launchd timer can all implement it, and the definition stays reviewable in the repo instead of buried in a scheduler config.
• Operations360 recorded as the long-term home for periodic jobs like this — the operations sibling to CINO-Customer-360 and Work360-*. No repo exists as of 2026-09-01, so it is written as a plan, not a dependency.
• Header and CLAUDE.md pointer corrected — both implied automation that does not exist. They now say manual, weekly, needs CSA-Internal gh.

Why the cloud path is parked rather than fixed

It needs two things: an environment setup step installing gh, and a PAT with CSA-Internal read access stored in the routine config. The second is a credential decision. The sweep's entire purpose is reading private orgs, and a local run already has that access without minting and storing a new org-scoped token — so the cloud route buys "runs when the Mac is off" at the cost of a new long-lived secret. That trade is worth making deliberately, if at all.

The one requirement carried forward

Whatever eventually runs this must keep exit 2 distinct from exit 0. "Could not check" and "nothing to check" are different answers; a green light nobody earned is more dangerous than no light. That is now stated as a spec requirement rather than an implementation detail of one script.

Verification

./tools/check-all.sh — all checks pass. Docs-only change; no script behaviour touched.