feat(app): persist dismissed provider subagents across app restarts

[SezginKahraman]

Linked issue

Discussion: #4440 (this is a feature request, not a bug — no issue to close). Opening as a draft per the discussion until there's agreement this is the right shape; see CONTRIBUTING.md guidance on unsolicited PRs.

Type of change

• Bug fix
• New feature
• Enhancement
• Refactor
• Docs

Reasoning

"Archive finished" hides completed/failed/canceled provider-owned (Task-tool) subagent rows only in an in-memory hiddenFromTrack Set (packages/app/src/subagents/provider-store.ts), documented as intentional in docs/agent-lifecycle.md. Every reload or app restart wipes it, since refreshProviderSubagents refetches the full descriptor list from the daemon on mount and the daemon has no concept of a "dismissed" provider subagent to filter back out. On a long-running parent that fans out a lot of subagents, this means re-dismissing the same finished rows over and over across reconnects. See #4440 for the full framing.

This PR persists just the hiddenFromTrack set (not the much larger, non-serializable-safe descriptors/timelines maps) to the device's local storage, using the same persist + createValidatedPersistStorage + AsyncStorage pattern already used by sidebar-collapsed-sections-store and workspace-service-route-preferences. The store construction moves into an exported createProviderSubagentStore(storage) factory (matching workspace-service-routes/store.ts) so tests can inject an in-memory storage and exercise a real write → restart → read round trip.

Goals

• A dismissal made via "Archive finished" for a provider-owned row survives a reload/app restart on the same device.
• No change to managed Paseo subagent archiving (already daemon-side and correct).
• No change to descriptors/timelines persistence — those stay in-memory only, refetched/rebuilt from the daemon as before.

Non-goals

• Not adding a daemon-side "archived" concept for provider subagents — this stays a client-local presentation dismissal, exactly as documented today, just a durable one. It does not propagate to other clients, and does not survive the provider reporting the same subagent id through a different daemon.
• Not pruning hiddenFromTrack against the live provider list — an existing test (keeps hidden state when a child temporarily disappears from the provider list) already asserts a dismissal must survive a subagent temporarily vanishing from a replaceList response, so the set isn't safe to prune on absence; it only grows. Flagged explicitly in the doc update. A size-bounded eviction policy would need a design decision I didn't want to bundle into this PR.

QA

Commands run:

npm run build:client && npm run build:server   # workspace declarations must be current, see CLAUDE.md
npx vitest run packages/app/src/subagents/provider-store.test.ts packages/app/src/subagents/select.test.ts packages/app/src/subagents/hidden-track-persistence.test.ts
# 28 passed, 0 unhandled rejections
npm run typecheck   # clean across all workspaces
npm run lint -- packages/app/src/subagents/provider-store.ts packages/app/src/subagents/provider-store.test.ts packages/app/src/subagents/hidden-track-persistence.ts packages/app/src/subagents/hidden-track-persistence.test.ts
# 0 warnings/errors

New test survives an app restart instead of resetting Archive finished dismissals (provider-store.test.ts) creates two separate store instances sharing one in-memory storage backend — the second simulates the app restarting — and asserts the dismissal from the first instance is visible in the second after persist.rehydrate(). Confirmed it fails (createProviderSubagentStore is not a function) against the pre-change code and passes after.

One thing I had to work around: @react-native-async-storage/async-storage's web build reads window.localStorage unconditionally, which crashed as an unhandled rejection when the existing provider-store.test.ts/select.test.ts tests (which mutate the real exported singleton, not an injected instance) ran in this package's Node-environment unit test project. resolveHiddenTrackStorage swaps in a no-op storage specifically for that gap (web-targeted evaluation with no window — native and real-browser evaluation are unaffected); covered by hidden-track-persistence.test.ts.

I did not test this manually against a running app (web only, per my role in the linked discussion) — the QA here is the automated round-trip test above, which exercises the real persistence mechanism rather than a mock.

Checklist

• One focused change
• npm run typecheck passes
• npm run lint passes
• npm run format passes
• QA evidence
• Tests added or updated where it made sense

[boudra]

Closing this feature PR. Please describe the workflow and shared need in Discussions, as described in CONTRIBUTING.md.