fix(cron): serialize lifecycle script reads with SQLite locks

[JoaoMarcos44]

Fixes #102589.

Root cause

cron/lifecycle_guard._read_referenced_script() raw-opened and closed every path-like token in the gateway process. When the token was the live state.db, the close could cancel SQLite's process-wide POSIX advisory locks and leave the WAL unprotected. A standalone has_live_connection() check would still leave a check/use race.

Fix

Use the existing hermes_cli.sqlite_safe_read.offline_file_access lifecycle lock around the complete open/read/close operation. A live tracked SQLite connection now causes the guard to fail closed without opening the file; normal script reads retain their existing behavior. This fixes the race class rather than adding a duplicate registry or parallel reader.

Verification

• scripts/run_tests.sh tests/cron/test_lifecycle_guard_live_db.py -q — 3 passed
• scripts/run_tests.sh tests/hermes_cli/test_gateway_restart_loop.py -k test_read_referenced_script_tolerates_nul_in_path -q — 1 passed
• ruff check cron/lifecycle_guard.py tests/cron/test_lifecycle_guard_live_db.py — passed
• git diff --check — passed
• Regression test uses a real connect_tracked() SQLite connection and proves os.open is not reached.

The requested infographic is generated locally at issue-102589-infographic.png and is intentionally not committed because repository CI rejects committed infographics.

[alt-glitch]

This was generated by AI during triage.

Related: #102596 is an earlier open fix for the same issue (#102589) using the sqlite live-connection registry; this PR uses the existing offline_file_access lock instead. Flagging so a maintainer can pick one approach.

[andrexibiza]

Reviewed exact head 1b2365d441c5b0f45063d7435e97d5970e587c4c against current main@63279301bcbdc185c1b07b98a9312eb0c862f26d.

No blocker found on this object. This is the right repair boundary for #102589: it reuses the existing SQLite lifecycle authority and holds that authority across the complete raw open/read/close sequence, so connection-open/close transitions cannot race a separate pre-check. A live tracked alias fails closed through offline_file_access; ordinary script reads and invalid-path tolerance stay intact. The PR is one commit, two files, directly based on current main, does not introduce a parallel registry, and the modified production owner remains below the repository's 2K ceiling.

Exact-head acceptance is clean: CI 33838652115, Docker 33838651598, and Nix 33838651605 are all green, and with one surviving commit the every-commit gate is satisfied. This should be the semantic carrier for the lifecycle-guard side of #102589; preserve any unique regression value/provenance from #102596 rather than landing two competing guards.

[crazyief]

Lab: tests/cron/test_lifecycle_guard_live_db.py → 3 passed. Covering the whole open/read/close under sqlite locks kills the race. Not an approval.

[teknium1]

Landed on main in #108076 (902ecca). Your commit was cherry-picked with authorship intact; thank you @JoaoMarcos44.