refactor(t096): arch + reliability hardening sweep (15 verified fixes)

[duongdev]

Investigation + refactoring sweep from /predict-issues + /improve-codebase-architecture, then adversarial verification of every finding against the real code. Verification refuted 17/20 predicted issues and 4/7 deepening candidates as already-mitigated (existing caps, watchdogs, guards) — this PR lands the genuinely-actionable remainder: 15 fixes, all evidence-checked.

Task: docs/tasks/096-investigate-for-arch-and-refactoring-predict-issue.md
ADR: docs/adr/0015-prefer-thin-handlers-over-reducer-indirection.md

What's in it

Reliability hardening

• Side-channel cdpCall timeout + reject-on-close; reap of hung CONNECTING sockets (P4/P3)
• Settings writeFileSync guarded in both backends (P15)
• Per-send push badge recomputed from a fresh list, not a stale snapshot (P7)
• Adaptive paint-ack watchdog window (EWMA of paint latency, floor 1s/cap 5s) so a slow device isn't tripped early (P2)
• Closed-tab stack capped at 50 (P18); Slack sweep triggers debounced per-workspace (A6)
• Electron onSwipe now returns an unsubscribe — fixes a per-reconnect listener leak (P6)
• find() rejection caught; command-palette run() guarded with a failure toast (P11/P19)

Architecture / locality

• Extracted createWsChannel / createInputChannel / createReconnectDriver out of the 1419-line cdp-web-transport.ts (now 970) into own files — WS channel went from 0 to 8 isolated tests (A5)
• Notification handlers route through one optimistic-mutate helper that reverts on a failed POST (A2)
• applyNextActive dedupes the close/switch tail; phoneView + readerEntry folded into one discriminated union (A1/A7)
• Settings dialog's dual getUiState load merged to one (A3); full useSettings consolidation spilled to t097 (needs visual review)

Per ADR-0015, the reducer/event-bus "deepening" proposals (A1/A2/A4/A6) were deliberately not built — the orchestration was already concentrated behind a seam, so small helpers were the right call.

Testing

Automated gates all green: pnpm typecheck, pnpm test (936), pnpm test:e2e (43), pnpm build, node --check on all 3 backends, pnpm check:changed (0 errors). New TDD coverage for closed-tabs cap, sweep-scheduler, paint-ack-pacer, side-channel reap/timeout, ws-channel, input-channel.

HITL pending (needs the iPad / live Remote Browser / Electron app)

• Layer-2 smoke: P6 (swipe×reconnect), P15 (unwritable settings), P4 (stalled cred socket), P3 (hung side-channel reap), P7 (push-while-marking badge)
• Layer-3 visual: A7 (phone reader nav), A3 (settings cards persist), P19 (palette toast)
• App-boot check (preload.js / main.js)

Opening for the per-branch preview deploy to run these on-device.

[dokploy-2026-03-01-6ei3s7]

Dokploy Preview Deployment

Name	Status	Preview	Updated (UTC)
cdp-browser-app	✅ Done	Preview URL	2026-06-20T08:08:14.034Z