release: v4.7.1 — liveness alarm actually opens issues (checkout + threshold)

[askalf]

What does this PR do?

Closes two latent bugs in the v4.4.2 liveness alarm that overnight observation exposed. The alarm had been firing every 2 hours and correctly detecting that the class-B watcher was past threshold — but failing before it could open the `cc-watcher-liveness` issue. Net result: the watcher-of-the-watcher was a no-op for the entire window from v4.4.2 (2026-05-17) through v4.7.0. If the class-B watcher had actually gone offline, no alert would have surfaced.

Bug 1 — missing `actions/checkout`

The workflow shelled out to `gh issue list` / `gh issue create` without first checking out the repo. `gh` resolves the target repository by reading `.git/config` from the working directory; without a git context, it exits with `fatal: not a git repository`. The workflow logged `Last successful watcher run: 2026-05-18T06:35:47Z (4 hours ago, threshold 3h)` and then crashed.

Fix: `actions/checkout@v6.0.2` at the top of the job. We don't actually need any files — just the `.git` directory `gh` needs.

Bug 2 — threshold set against fictional cadence

I sized the v4.4.2 threshold (3h) against the declared `*/30 * * * *` cron (= 6 missed cycles). But GitHub Actions' free-tier cron scheduler is best-effort, not guaranteed. The observed cadence of the class-B watcher on this repo overnight was every 2-4 hours:

```
23:46:53 → 02:08:34 (2h22m)
02:08:34 → 06:35:47 (4h27m)
06:35:47 → 10:56:08 (4h21m)
```

So healthy watcher state would trip the 3h threshold ~half the time — but never actually surface as an alert because of Bug 1 above. Two bugs masking each other.

Fix: `THRESHOLD_HOURS` 3 → 8. Absorbs the observed skew while still catching real outages (anything past 8h of silence is signal, not noise).

Alert body text updated

The alert body now describes both the declared and observed cadence so an investigator reading the issue understands the threshold rationale instead of asking "why 8 hours?"

Documented

`docs/drift-monitor.md` gains an explicit "Observed cadence" column distinguishing declared from real-world cron. Plus a paragraph stating: GitHub Actions free-tier cron is best-effort; if you need sub-hour SLA, self-host the runner and the cron driver (e.g. a cron entry on the same Hetzner box calling `gh workflow run cc-drift-template-watch.yml --ref master`).

How to test

```bash
git fetch origin fix/v4.7.1-liveness-alarm
git checkout fix/v4.7.1-liveness-alarm
npm run build && npm test # 75/75 (no src/ changes)

Manual trigger after merge — should now exit 0 if watcher last

succeeded within 8h, or actually open a labeled issue if past:

gh workflow run cc-drift-watcher-liveness.yml --ref master
```

Checklist

• `npm run build` passes
• `npm test` passes (offline regression test, no credentials required) — 75/75
• For changes that touch `proxy.ts`, `cc-template.ts`, or streaming behavior: tested with `dario proxy --verbose` + `node test/compat.mjs` (requires credentials) — N/A: workflow + docs only
• No new runtime dependencies added
• No tokens/secrets in code or logs