Problem

Desktop session persistence used one deterministic temp path per session (<sid>.json.tmp) while releasing self.lock before disk I/O. Two threads persisting the same session could therefore write/replace the same temp file concurrently. A separate deletion race also allowed a late worker persist to recreate a session JSON after the session had been deleted from manager state.

Root cause

_persist_session() only held self.lock while building the data snapshot, then performed tmp.write_text() and os.replace() unlocked. _delete_session_file() also performed its filesystem operation without the same lock, and _persist_session() did not verify that the Session object was still registered before writing.

Fix

• keep the existing RLock held across the session snapshot, temp write, and atomic replace
• skip persistence when self.sessions.get(s.id) is not s, preventing stale workers from recreating deleted sessions
• perform session-file deletion under the same lock

This preserves the existing per-session JSON format and atomic replace strategy.

Regression coverage

• concurrent persists of the same session cannot overlap writes to <sid>.json.tmp
• a late persist of a deleted Session object cannot recreate its JSON file
• the test harness loads desktop_bridge.py only inside the unittest class lifecycle, uses an isolated temporary GA_ROOT, restores sys.argv and sys.modules, and does not leave import-time global state behind

Verification

TDD was performed on a separate validation branch:

1. Regression-only GitHub Actions run 31164526542 failed both targeted cases on current main
   • concurrent same-session temp writes overlapped and produced an actual FileNotFoundError from os.replace
   • a late _persist_session() recreated the file after delete_session()
2. The minimal locking/registration fix passed run 31164598932: compile, both deterministic concurrency regressions, and git diff --check
3. Review feedback identified test-global-state leakage. The harness was refactored without changing production code, and validation run 31165250013 passed compile, both regressions, and git diff --check
4. The contribution branch was then rebuilt directly from current main as one clean commit containing the production fix plus the isolated regression test

Scope

• 1 production file modified
• 1 focused concurrency regression-test file added
• 1 clean commit on top of current main
• no dependencies or runtime configuration changes