[fix] - post-security-audit robustness + accuracy follow-ups (4 small nits)

[cgfixit]

Proposed changes

Four low-severity follow-ups from the Phase 2 security scan (the audit found no high-severity issues; the codebase is heavily hardened). Each traces to a specific finding and is independently trivial:

1. Temp-file orphan on write failure (utils/ops_runner.py _write_body): the file is created (NamedTemporaryFile(delete=False)) before handle.write(body); run_agentic_op only unlinks the name _write_body returns, so a mid-write failure (disk full, un-encodable body) orphaned a cyclaw_skill_*.md in the OS temp dir. Now cleans up its own orphan on the write-error path (close-then-unlink, Windows-safe). Regression test forces the write to raise and asserts no orphan survives.
2. Listing race (harness/sessions.py SessionStore.list()): sorted by getmtime before the per-file try/except SessionStoreError skip loop, so a session file removed between glob and sort raised OSError straight out of a path meant to tolerate missing/corrupt files — 500-ing /api/sessions + /api/status. Sort now uses a getmtime helper returning 0.0 on OSError. Regression test races a vanishing file.
3. Dead branch (llm/client.py): if not _is_loopback_url(primary_url): pass computed a loopback check and did nothing — reads like enforcement but is a no-op. Replaced with a plain comment (the non-loopback-primary allowance is intentional).
4. Docstring accuracy (utils/logger.py): the module docstring claimed query hashing is unconditional, but it's gated on logging.audit_fields.include_query_hash (shipped true; the scoped guarantee is already pinned in test_due_diligence_invariants.py). Docstring now states the toggle and that disabling it is privacy-affecting — doc fixed to match tested behavior, not a behavior change.

Invariant / Governance Impact: None — no graph edge, gate auth, sanitizer, or soul path touched. invariant-guard 28/0. #4/#3 are the security-relevant ones and neither weakens a control (temp-file hygiene; a doc that now correctly describes an already-tested privacy toggle).

Types of changes

• Bugfix (non-breaking change which fixes an issue)

Scope note: Out-of-band + support code (ops_runner, harness, llm client, logger docstring).

Benefits / why

• Two real (if low-severity) robustness bugs on error/race paths get regression coverage; two cosmetic/accuracy nits removed. Net: fewer ways for an operator's out-of-band tooling to leak a temp file or 500 a listing, and an audit docstring that no longer overpromises.

Risks to monitor

• _write_body's except catches (OSError, UnicodeError) — the realistic write-failure modes; a KeyboardInterrupt mid-write still won't clean up (accepted, matches the surrounding code's exception discipline).

Further comments

Verified: ruff clean on all 6 files; 104 passed (test_ops_runner + test_harness + test_audit); both regression tests revert-proven (fail against origin/main source, pass on the fix); invariant-guard 28/0. Shares only tests/test_harness.py with the open #651 (both append tests, non-adjacent) — trial-merge verified clean, noted below.

🤖 Generated with Claude Code

https://claude.ai/code/session_0138hmScu3tFU4v1jzvFUwVT

---

Generated by Claude Code