daily-fix: subfloor reclaim never fired + local-date logs

Workflow gap

Two watcher defects found during the 2026-08-04/05 disk crisis:

1. Subfloor reclaim silent no-op. With / under the 60 GiB arm point (EPM_VM_DISK_SUBFLOOR_GIB, default 60 — probed at autonomous_session_watch.py:4429) for ~7 hours, "the subfloor auto-reclaim has never fired. Zero subfloor lines all day … And the guard logs vm-disk: ok (42.2 GiB free)." Kill switch NOT set (probed at mining time). Manual reclaim was the only thing that recovered space — the #658-class full-disk safety net did not engage.
2. Log naming defeats UTC greps. The watcher's daily log is named by LOCAL date; the PM session grepped $(date -u +%F).log and read a nonexistent file for the 7 hours local lags UTC — its "no subfloor lines" reads were empty for the wrong reason before a corrected re-check reached the same conclusion. Same-session sibling of an earlier completion-cutoff timezone bug.

verified-at-filing: both are the recovery miner's probed reads (session 4966e56e rows 1244–1255; source probe of the arm-point constant + settings kill-switch absence at mining time). unverified hypothesis — verify at plan time: the non-firing mechanism is inside the pass predicates (rate-limit/single-flight/dry-run state) — the tick calls both passes every cycle, so the gate is predicate-side; not reproduced against a <60 GiB fixture.

Proposed change

In scripts/autonomous_session_watch.py: (1) audit subfloor_reclaim_pass gating against a below-floor fixture and make a non-firing loud (a sidecar row per skip naming which predicate blocked); (2) name watcher daily logs by UTC date or symlink both forms.

Provenance

• fingerprint: 94421c23b244

• workflow_fix_target: scripts/autonomous_session_watch.py

• origin: /daily 2026-08-04 recovery sweep — miner 7 P9/P11.