fix(desktop): guard session.info state patches with explicitSid

[yingliang-zhang]

Summary

The desktop event handler in gateway-event.ts resolves sessionId = explicitSid || activeSessionIdRef.current. When explicitSid is empty (unscoped event), events fall back to the active session — which is correct for display (deltas, completions are the focused turn's own output, see #42178). But the session.info state-patch path (updateSessionState) used the same fallback sessionId to patch per-session model/cwd/branch.

During a session switch, an unscoped session.info could patch the active session's state attributes with a background session's values — the Race 3 gap left open by #53936.

Fix: require explicitSid for the updateSessionState call in the session.info handler. Display operations (delta append, completion) still use the fallback — only per-session state mutations are gated.

Before (leaks state patches via fallback sessionId)

if (sessionId && hasStatePatch) {
  updateSessionState(sessionId, state => ({ ...state, ...statePatch, ... }))
}

After (only patches when we have a real session_id)

if (explicitSid && sessionId && hasStatePatch) {
  updateSessionState(sessionId, state => ({ ...state, ...statePatch, ... }))
}

Why this is safe

• Background sessions with explicit session_id still get their state patches (the common case).
• Unscoped session.info events from the active turn still apply global setters (setCurrentModel, setCurrentCwd, etc.) via the existing apply guard — only the per-session updateSessionState call is gated.
• Display operations (message.delta, message.complete, reasoning.delta, tool.start, etc.) are unaffected — they use sessionId for rendering but don't call updateSessionState with state patches.

Relationship to existing work

PR	Scope	Status
#53936	TUI-layer null-sid + empty-string session_id filter (Races 2, 2b)	Open
This PR	Desktop Electron layer session.info state-patch guard (Race 3)	—

#53936 explicitly notes the desktop Electron layer fallback as remaining. This PR closes that gap.

Tests

• NODE_ENV=test npm --workspace apps/desktop run typecheck — passes
• vitest run src/lib/gateway-events.test.ts — 3/3 pass
• NODE_ENV=production npm --workspace apps/desktop run build — passes

Checklist

• No personal info / internal paths in diff
• Git author is zhangyingliang@outlook.com
• No secrets / .env in diff

[teknium1]

Thanks for isolating the state-patch path rather than changing display-event routing.

Problems

• The changed guard has no direct regression coverage. apps/desktop/src/lib/gateway-events.test.ts:21-25 only checks the classifier; it does not invoke useGatewayEventHandler or verify updateSessionState for an unscoped, state-patched session.info.
• The claimed unscoped background-event route was not established from current main. _emit always serializes session_id in tui_gateway/server.py:1143-1147, while the desktop routing contract says missing IDs represent the focused turn in apps/desktop/src/lib/gateway-events.ts:21-23.

Suggested changes

• Add a hook-level regression test for an unscoped state patch with an active session (no per-session mutation), plus an explicit-ID event that updates the addressed session.
• Identify the valid gateway emission path that produces the Race 3 event so the new session.info exception is covered by a concrete contract.

Automated hermes-sweeper review.

[teknium1]

Please add a hook-level regression test for this guard. The current gateway-events.test.ts only permits unscoped session.info; it does not verify that an unscoped state patch cannot call updateSessionState for the active session, while an explicit session ID still can.

[yingliang-zhang]

Closing after re-auditing the current gateway/Desktop contract against the maintainer review.

On current main, _emit() always serializes session_id; WebSocket/shared routing forwards params unchanged; all identified background session.info producers supply explicit IDs. Missing IDs are explicitly documented as focused-turn output, not background state. The queue source-session changes from the old branch are now patch-equivalent upstream. Therefore no valid unscoped background state-patch producer remains, and adding a synthetic hook test would preserve speculative behavior rather than a production contract.

Verification on current main: Desktop gateway-events Vitest 6 passed, Desktop typecheck PASS, and the replacement worktree has no delta. Thanks for the review—it correctly exposed that the original Race 3 premise was not established.

[yingliang-zhang]

Correction to my previous comment: the old combined commit is not wholly patch-equivalent to current main (git cherry origin/main still reports it unique). The close decision remains correct for this PR's stated Race 3 scope because the unscoped background session.info producer path does not exist. The separate queued-prompt source-boundary behavior from that historical combined branch is being handled in the broader #63298/local integration line, not treated as already merged here. I should not have described that queue portion as upstream-equivalent.