fix(ops): allowlist deliberately-absent llm-worker in compose-drift sweep (#3129)

[Xore]

Why

llm-worker is operator-gated down per #3023 for the #1947 cold benchmark run. The #3040 rule correctly alarms on a missing expected-persistent service even with no siblings to compare against — that's the exact shape that let llm-worker sit invisible for 11h. But #3040 gave no way to mark a known, reasoned absence, so it re-fires on the same intentional state (this PR closes #3129's re-fire).

Adds EXPECTED_ABSENT_WHILE allowlist (name → reason string citing #3023), checked in sweep() Gate 1. This implements the mechanism #3040 itself asked for, not a blanket sibling-suppression — every other missing-service case still alarms.

Evidence

• scripts/compose-drift-watch.py (+26/-0): allowlist + Gate 1 check, comment cites ops: nothing detects a declared-but-absent container, so a stack can lose exactly one service in a rebuild unnoticed #3040 authorization.
• scripts/tests/test_compose_drift_watch_sweep.py (+34/-3): 3 new tests — llm-worker quiet, ml-worker still alarms as control, general pin re-pointed at auth-events-worker.
• 25/25 tests pass.

Refs #3129

🤖 Generated with Claude Code

[strix-security]

Strix Security Review

No security issues found.

_{Updated for dcf56f6.}

---

Reviewed by Strix
_{Re-run review · Configure security review settings}

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[strix-security]

Reviewed the two changed files in full. The PR adds an EXPECTED_ABSENT_WHILE allowlist (keyed by project name with a documented reason) to the internal compose-drift-watch.py monitoring script, plus a Gate 1 check that skips alarming for the single named llm-worker project, and three corresponding unit tests. The allowlist key is a hardcoded string and the check performs a plain membership test on a directory name; no attacker-controlled input reaches any dangerous sink, and there are no secrets, auth, crypto, or injection changes. Bandit and semgrep found no issues in the changed lines (Bandit's low-severity subprocess notes are all pre-existing code outside the diff). The alarm-suppression is an explicitly documented, temporary operator decision and does not weaken access controls or expose data. No security vulnerabilities found.

---

Reviewed by Strix
_{Configure security review settings}