docs(runbook): self-heal can loop forever + R8 v2-primary cutover traps

[MaxGhenis]

Two operator lessons from the live macOS install (2026-07-19) that weren't written down anywhere — both cost real time. Docs-only, no code.

Google self-heal can loop silently forever

The #74 in-process cookie refresh assumes it can renew from Chrome. On an install whose Chrome has zero messages.google.com cookies (never opened Messages-for-web), there's no source to refresh from, so /api/status sits at needs_repair:true / "…session cookie expired; refreshing and reconnecting…" indefinitely. A fresh re-pair reconnects but only holds ~3–4 min before the same expiry — so connected:true right after a re-pair is not proof of a fix (watch ≥5 min). Added a callout under the self-heal section with the sqlite3 check for the 0-cookie case and pointers to #108 (symptom) and #75 (durable Android-companion fix).

R8 / v2-primary cutover operator traps

New section. The V2Primary defaults lever (#142) has four non-obvious ways to silently no-op on macOS: a stale sandbox container hijacks defaults write (use CFPreferencesSetAppValue + verify against the running backend's env), killall cfprefsd can drop an unflushed write, build.sh's find .build -name OpenMessage first-match is nondeterministic with stale caches (clean-build first), and short Swift string literals like V2Primary are invisible to strings (grep the ≥16-byte env-var name to verify a build).

Not urgent — safe to merge whenever; documents things the next operator (or the in-flight refactor) will otherwise rediscover.

[MaxGhenis]

Closing as stale (idle since 2026-07-20; #148 fixed the self-heal loop it warns about by dropping the OSID requirement, and main's runbook now documents that fix, so this callout would add outdated advice; the stale-build checks are covered by #153 and #154). Max's PR sweep 2026-09-25. Reopen if still wanted.