[iamnbutler]

Follow-up to the #75 live-check gaps (items 1 and 2), plus a docs/live-checks.md refresh.

Isolated Chrome. Session.launchChrome() launches Google Chrome from /Applications or ~/Applications with a new temporary --user-data-dir, --no-first-run, --no-default-browser-check and --new-window, opening the fixture page. It finds that pid's target through desktop_apps and waits for the fixture window. quitChrome() runs desktop_quit on that target only. If the process is still running after 8 s, it sends SIGTERM to that child pid only, then deletes the profile. Each Chromium check gets its own launch and quits it in a finally; the global cleanup also covers Control-C. When Chrome isn't installed, the checks skip with a clear message instead of failing. The checks no longer skip while the user's own Chrome is running, and never touch its windows.

Checks

• feat: turn on Chromium and Electron web accessibility before inspect and read, once per app launch #107: inspect on a fresh launch reports web_accessibility: enabled and lists the controlled field, then already_on, then nothing when opted out. The assertions are unchanged; only the launch changed.
• feat: turn on web accessibility for desktop_wait's element and text polls too #110: desktop_wait element_appears for an AXTable (only the page has one) matches on a fresh launch with enabled. A text_contains wait with web_accessibility: false still matches and reports nothing.
• feat: retry an incomplete Accessibility read once in inspect and read, and say retried when it fails again #111 (new): reads the page the moment it opens. It passes on page text, or on a failure with retried: true, and fails on a failure without retried.

Check sections are now grouped by the release that shipped them (0.13.0 / 0.12.0 / 0.11.0); #114 and #99 were sitting under 0.10.0.

Docs: the Chrome isolation, the #114 Finder paste fixtures, what #111 covers versus a forced incomplete read (still manual), and that Pause is user-only and never driven by the harness. #100 Pause and the Electron/activity checks stay manual.

No CHANGELOG entry, since this is a dev-harness change.

Verification

• npm test passes 70/70 and npm run check is clean. node scripts/live-check.mjs --list shows the new order.
• --only 107,110,111 refuses on the locked screen, exits 2 and launches nothing.
• Offline probe of the launch and cleanup path, run without the MCP:
  • the spawned pid stays the browser process and doesn't hand off to the user's Chrome;
  • the temporary profile fills in;
  • SIGTERM exits it cleanly with no helper processes left;
  • the profile is removed;
  • the user's Chrome processes are unchanged.
• Live run: pending: screen locked.