fix(peer): improve directory picking and recover transient disconnects

[bobleer]

Summary

Fix Peer Device Mode directory picking and recovery from transient connectivity failures.

• Start the project directory picker in the controlled host's user home and keep the address field permanently editable at a consistent width. Enter navigates; Escape restores the current path.
• Keep the selected peer surface and adapter during presence gaps or RPC outages. Show a reconnecting notice with a manual return-to-local action, retry with backoff capped at 15 seconds, and restore healthy status only after the event attachment is acknowledged.
• Separate ordinary command failures from connection health accounting and serialize recovery probes so concurrent errors do not trigger competing reconnects.

Type and Areas

Type: bug fix / UI/UX

Areas: Web UI, Peer Device Mode, shared system service, desktop/Tauri and CLI system-info adapters, locales, architecture documentation.

Motivation / Impact

Opening a project on a peer previously started in an existing workspace or a filesystem root, and clicking its address replaced the full-width display with a short input. The picker now uses the peer's home directory and one stable input.

A brief absence from relay presence, or several failed product RPCs, could previously switch the UI back to the local device. Transient failures now retain the remote context while recovery continues; explicit disconnect, switching device, and logout retain their existing behavior.

Verification

Passed focused checks:

pnpm --dir src/web-ui run test:run src/infrastructure/peer-device/PeerDirectoryBrowser.test.tsx src/infrastructure/peer-device/PeerDirectoryPickerHost.test.tsx src/infrastructure/peer-device/peerDirectoryPath.test.ts
pnpm --dir src/web-ui run test:run src/infrastructure/peer-device/PeerConnectionManager.test.ts src/infrastructure/peer-device/PeerDeviceSurfaceController.test.ts src/infrastructure/peer-device/PeerConnectionStatus.test.tsx src/infrastructure/api/adapters/peer-device-adapter.test.ts
pnpm --dir src/web-ui run test:run src/app/services/sessionSceneLifecycle.test.ts src/flow_chat/services/sessionActivation.test.ts src/flow_chat/services/storeSync.test.ts src/app/stores/sceneStore.test.ts src/flow_chat/services/flow-chat-manager/PeerSessionRefreshModule.test.ts
cargo test -p openbitfun-services-core --no-default-features --features process-runtime --lib system::info::tests
cargo test -p openbitfun-desktop --lib system_info_home_contract
cargo test -p openbitfun-cli --bin openbitfun system_info_home_contract
pnpm run fmt:rs
pnpm run check:web
pnpm --dir src/web-ui exec eslint src/infrastructure/peer-device/PeerConnectionManager.ts
pnpm run i18n:audit
git diff --check upstream/main...HEAD
node scripts/check-git-object-sizes.mjs --base upstream/main --head HEAD

• Directory picker: 17 frontend tests and 4 shared-service/desktop/CLI tests passed, including legacy system-info payload compatibility.
• Recovery: 92 manager/controller/status/adapter tests and 61 session lifecycle/refresh tests passed. Coverage includes a 2.3-second presence gap, outages beyond the former failure threshold, concurrent command failures, event-attach ordering, stale callbacks, explicit disconnect, and logout.
• Browser QA with the real components and simulated peer RPCs confirmed unchanged address-field dimensions during editing and retained peer context/draft through outage and recovery. No screenshots are attached.
• The contributor reported successful manual testing of the pushed branch before requesting this PR. Agent-run remote verification used simulated Peer Device Mode; no live two-device test was performed by the agent. Remote workspace, mobile/bot remote control, and detached dispatch were not exercised.
• Broad builds and platform matrices are left to CI.
• The first CI run found a callback forward-reference lint error. The callback now checks adapter identity to retain stale-connection isolation; focused ESLint, all 25 connection-manager tests, and check:web passed after the fix.

Reviewer Notes

The shared system service supplies an optional home directory through existing desktop and CLI get_system_info responses. Older peers without this field get an explicit unavailable hint and can still navigate by entering a path; the picker never falls back to the controller's filesystem. Recovery uses the existing ping and attach protocol without new wire requirements or persisted state changes.

AI-assisted contribution. Testing level: fully tested for the focused behavior described above; cross-platform and other remote-scenario coverage is limited as noted.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.