feat(watch-pr): hold READY until review activity settles

[chhoumann]

Why

The commonest babysitting failure in an audit of agent history (about 121 of 227 flaw instances) was calling a PR ready once CI turned green, before review bots had posted. watch-pr knew only Bugbot. This adds a review settle window that names no bot.

What changed

• watch-pr --settle <seconds> (default 300, 0 off) holds READY until accounts other than the PR author and the gh viewer have been quiet on the PR for that long. The window starts at the latest of the head's push, the PR's opening, and its last ready for review. An eyes reaction holds READY up to 30 minutes after that start.
• The push time comes from GitHub's repository activity log, not commit dates, which a rebase keeps. While the window is open, the watcher waits as review-settling, and --status-only shows the time left.
• t3-tools.md says the standalone babysit skill's own polling is replaced on T3, and that no per-bot rules are needed.

Scope

• Covers single and --stack watches. Queued mode does not settle.
• Leaves out a cap on the activity hold and a window that restarts when CI turns green. Both are open questions for the maintainer.

Blast Radius

Every READY from watch-pr now waits at least until the window ends. The draft, review-required, and branch-protection gates also wait for it. The logic lives in the new port-only settle.ts. The upstream watcher files carry small hooks, listed in the 1.0.5 CHANGES entry for the next pstack-claude merge.

Verification

• The watcher suite passes (208 tests, 22 of them new), and 16 targeted mutations each fail the test aimed at them. Repo tests (1202), generate --check, the fork dry-run, and the sheet checker pass.
• The real reader on feat(cli): name the current note explicitly with current= and report currentNote quickadd#2196 put the window's end at 09:29:53, 5 minutes after Cloudflare's last comment edit. That PR was merged at 09:29:04, inside the window.