Verification review — reviewed the full diff (2 files, +157/-1).

Clean port of the EXDEV/EBUSY fallback from gemini-cli#21541:

1. Error scoping — only EXDEV (cross-device) and EBUSY (bind-mount busy) trigger the fallback; other OSError codes propagate unchanged. This is the correct narrow scope.

2. Fallback sequence — shutil.copyfile → shutil.copymode → os.fsync → os.unlink. The fsync ensures data is durable before the temp file is removed. Permission copy failures are caught and ignored (correct — the file is still readable).

3. Symlink preservation — the fallback writes to real_path (the resolved target), so symlinked configs survive cross-device writes. This maintains the [Bug]: atomic writes to HERMES_HOME files replace symlinked targets (config.yaml/SOUL.md) #16743 invariant.

4. Test coverage — parameterized EXDEV/EBUSY, symlink preservation, other-OSError propagation, and a real cross-device E2E test using /dev/shm vs tmpdir.

No issues found. Clean utility fix.

Verification review — reviewed the full diff (2 files, +160/-0).

The assistant transcript backfill guard is well-scoped:

1. Tail-ID tracking — _db_tail_id() captures the last message ID before the turn starts. After the turn, _ensure_visible_response_persisted() only inspects messages newer than that tail, avoiding false positives from pre-existing transcript.

2. Duplicate prevention — the guard checks any(message.get("role") == "assistant" for message in new_messages) before writing. If the agent already persisted its response, the guard is a no-op. Correct dedup.

3. Failure mode — the entire guard is wrapped in try/except with logger.debug. A persistence failure is non-fatal (the response was already sent to the platform). This is the right degradation for a defensive backfill.

4. Test coverage — three cases: backfill when agent missed, no-op when agent persisted, no-op on empty response. The _FakeSessionDB mock correctly simulates the append + get_messages interface.

No issues found. Clean gateway resilience fix.