fix(hermes_state): ensure state.db created with 0o600 permissions, not 0o644 umask default

[liuhao1024]

What does this PR do?

Before this fix, state.db was created by sqlite3.connect() with the process umask (usually 0o022), landing at 0o644 (world-readable). On default HERMES_HOME=0700 this is inert (directory traversal blocks access), but it becomes load-bearing when HERMES_HOME_MODE is wider (e.g., 0o755 for web-server traversal use cases) or in managed mode (NixOS sets it to 0o750). In these scenarios, other local accounts on the same host can read the full conversation history stored in state.db.

This fix follows the TOCTOU-safe pattern already used for auth.json in hermes_cli/auth.py:

1. Pre-create state.db with 0o600: Before sqlite3.connect() touches the file, we create it with os.open(O_CREAT | O_EXCL, 0o600) to avoid umask-based permissions.
2. Handle race conditions: If the file already exists (e.g., created by another process between our exists() check and os.open()), we chmod it to 0o600.
3. Chmod existing files: If state.db already exists (e.g., after a repair or from a prior run), we ensure it has 0o600 permissions.
4. Secure WAL/SHM sidecars: After apply_wal_with_fallback() sets WAL mode, we chmod the -wal and -shm sidecars to 0o600, since SQLite creates them at the process umask and they hold recent uncommitted writes.

The fix is a security hardening measure: even if the parent directory has wider permissions, the database file itself is not world-readable.

Related Issue

Fixes #59706

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• 🔒 Security fix
• ✅ Tests (adding or improving test coverage)

Changes Made

• hermes_state.py: Pre-create state.db with 0o600 in SessionDB._connect_and_init() before sqlite3.connect(); chmod existing files; chmod -wal/-shm sidecars after WAL mode is applied.
• tests/test_hermes_state_permissions.py: Add regression tests for:
  • Fresh database creation with 0o600 under umask 0o022
  • Existing database being chmod to 0o600
  • Race condition handling (file created between exists() and os.open())

How to Test

1. Reproduce the bug (before the fix):

   import os, sqlite3, stat
   os.umask(0o022)
   conn = sqlite3.connect("/tmp/test-state.db")
   conn.execute("CREATE TABLE t (x int)")
   conn.commit()
   conn.close()
   print(oct(stat.S_IMODE(os.stat("/tmp/test-state.db").st_mode)))  # 0o644

2. Verify the fix:

   cd /tmp/hermes-bugfix-Mrc2jN
   python3 -m pytest tests/test_hermes_state_permissions.py -v

   Expected: All 3 tests pass, confirming:

   • state.db is created with 0o600 permissions
   • Existing databases are chmod to 0o600
   • Race conditions are handled correctly

3. Manual verification:

   import os, sqlite3, stat
   from pathlib import Path
   from hermes_state import SessionDB
   
   os.umask(0o022)
   db_path = Path("/tmp/test-sessiondb")
   db = SessionDB(db_path=db_path)
   
   mode = stat.S_IMODE(os.stat(db_path).st_mode)
   print(f"state.db mode: {oct(mode)}")  # Should be 0o600
   
   wal_mode = stat.S_IMODE(os.stat(str(db_path) + "-wal").st_mode)
   print(f"-wal mode: {oct(wal_mode)}")  # Should be 0o600
   
   shm_mode = stat.S_IMODE(os.stat(str(db_path) + "-shm").st_mode)
   print(f"-shm mode: {oct(shm_mode)}")  # Should be 0o600
   
   db.close()

   Observed result: All files show 0o600 permissions, not 0o644 (umask default).

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 15.2

Documentation & Housekeeping

• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — os.chmod is a no-op on Windows but the fix still provides defense-in-depth on Unix-like systems
• I've updated relevant documentation — or N/A (internal security hardening, no user-facing behavior change)

[alt-glitch]

This was generated by AI during triage.

Duplicate of #59716 (earliest-open twin, created ~2s earlier) — both fix issue #59706 at the same site (hermes_state.py SessionDB init) with the same mechanism: pre-create state.db via os.open(O_CREAT|O_EXCL, 0o600) before sqlite3.connect() and chmod the -wal/-shm sidecars to 0600. Related to issue #59706 and the 0600-perms family (#31469, #19699).

[liuhao1024]

Closing this PR as a duplicate of #59716 (earlier-open twin, created ~2s earlier). Both fix issue #59706 at the same site (hermes_state.py SessionDB.__init__), and #59716's approach aligns with the existing TOCTOU-safe pattern used for auth.json in hermes_cli/auth.py.