[AmrMsCLL]

What

Phase 7 of the hub plan: every route at 1440 and 390 in Chromium and Firefox, a keyboard-only pass on both editors, a reduced-motion pass, a load test of / and /live at 200 tabs and 50 events a second, a Redis-loss test and the backup cycle. The defects it found, fixed in four commits:

1. Reduced motion. Six state pulses (reconnecting dot, armed confirm, saving indicator, stage live dot, editor connection dot) looped regardless of prefers-reduced-motion. They now sit in one no-preference block beside the keyframes. With the preference set, document.getAnimations() reports nothing running on any public route in either browser.
2. Post editor from the keyboard. The formatting toolbar is one tab stop with Arrow, Home and End inside it instead of 22 stops; the cover upload is a button (the <label> around a hidden input was never focusable); the tags field keeps the typed text so a comma can be entered; the post page's heading list no longer picks up Tailwind's outline utility.
3. Serving through a Redis outage. A member's request answered 500 while Redis was down; it is now served as a visitor with the cookie kept. /health/ready stays 200 while SQLite and the snapshot work (the production healthcheck feeds Traefik) and says when the bus is down. Failed jobs back off 1, 2, 4… minutes up to an hour and give up after eight attempts instead of exhausting five attempts within a millisecond; a job that cancels its own row while running keeps the row. Publishing refuses while the bus is unreachable, so a scheduled post due during an outage is retried whole instead of being published without its event and its subscriber emails; the emails are queued before the event is recorded.
4. Redis command timeout of two seconds, so a page during an outage is served in about two seconds rather than 10 to 15.

Verified

• Firefox and Chromium, both widths, 24 public routes on the production build plus the signed-in routes in dev: every route 200, no overflow, no console errors.
• Load, production build locally: 200 streams open in 389 ms; 1,420 events published in 30 s, every tab received every event, 0 stream errors; GET / under the storm with eight concurrent loops p50 348 ms, p95 511 ms, p99 670 ms, 0 errors; RSS 389 MB with the streams open.
• Redis loss: pages serve as visitor and member, /health/ready 200, the scheduled post is published on the first retry after Redis returns, the subscriber email arrives, /blog lists it.
• Backup: a production run through the service, archive and inner checksums match, integrity_check ok, imported locally with members, profiles, projects and settings intact (production has no posts yet; posts were exercised on top of the import).
• npm run typecheck, npx eslint ., npx prettier --check ., npm run build, npm run size (largest page 136.9 KB). Specs run: server/jobs.test.ts, server/posts.test.ts, server/auth.test.ts.

A separate PR on platform fixes the backup service crashing on a malformed request. The full plan and results: .claude/.Plans/as2026-11-07-graphite-hub-phase-7.md.