fix: persistence safety — resume status, test isolation, write race

[aterrylu]

Summary

Three critical persistence fixes that caused data loss in production:

• Resume status bug: createSession() now explicitly sets status: "running" when calling persistSession(), so resumed sessions clear their stale "exited" status instead of preserving it via the ?? merge fallback
• Test isolation: All persistence tests now use isolated temp directories via _setConfigDirForTesting() instead of writing to production ~/.autonomos/sessions.json (which corrupted real agent data)
• Write race: Added in-memory cache to persisted.ts — eliminates TOCTOU races where concurrent markSessionExited() and persistSession() calls could overwrite each other. Added backup-on-shrink before destructive writes

Changes

File	Change
sessions.ts	Add status: "running" to persistSession() call
configDir.ts	Add getConfigDir() with test override support
persisted.ts	In-memory cache, dynamic file path, backup-on-shrink, defensive copy on getPersistedSessions()
backward-compat.test.ts	Use isolated temp directory
orgChart.test.ts	Use isolated temp directory
duplicate-name.test.ts	Use isolated temp directory for resume tests

Architecture

graph TD
    A[createSession] -->|status: running| B[persistSession]
    C[markSessionExited] --> D[readSessions - cached]
    B --> D
    D -->|first call| E[loadFromDisk]
    D -->|subsequent| F[return sessionsCache]
    B --> G[writeSessions]
    C --> G
    G -->|count decreased| H[copyFileSync backup]
    G --> I[writeFileSync]
    I --> J[sessionsCache = sessions]

Loading

Test plan

• All 85 tests pass (make check)
• Tests use isolated temp dirs — production sessions.json not modified
• status: "running" explicitly set on session creation/resume
• In-memory cache prevents concurrent write races
• Backup created before writes that reduce session count
• getPersistedSessions() returns defensive copy

🤖 Generated with Claude Code

[nox-0x]

Good, thorough fix. The in-memory cache is the right approach for the TOCTOU race, and explicitly setting status: "running" closes the resume bug cleanly. Test isolation to temp dirs is much safer — good call.