workflow-fix: test_tick_triage module-level NOW makes fixtures wall-clock-sensitive in long gate sessions

Target file: tests/test_tick_triage.py (line 37: NOW = time.time(); consumer _event(kind, age_s) writes ts = NOW - age_s).

Defect. NOW is captured once at module IMPORT (pytest collection, session start), while tick_triage.triage() measures marker/transcript ages against the LIVE clock at test EXECUTION. In any pytest session where the file executes ≳20 min after collection, every fixture event written as "fresh" (e.g. _event("epm:progress v1", 300) = 5 min old) is silently ~drift+5 min old at assertion time — old enough to cross tick_triage's generic staleness threshold, which then pre-empts the specific code path under test.

Incident (#2158 Step 9c gate, 2026-08-18). A 234-file gate run (1h34m under fleet load ~16) failed exactly 4 tests, all in the api-error-after-marker probe family: test_api_error_after_marker_returns_stale_redrive, test_api_error_after_marker_falls_open_on_missing_transcript, test_api_error_probe_kill_switch, test_api_error_probe_incident_1689_content_string. Observed failure reason: status=running, marker age 28m — bg poll chain likely dead — the junit shows 22.2 min of cumulative test time before the first tick_triage case, so the 5-min fixture marker was ~27–28 min old. Exact arithmetic match. The file is byte-identical to origin/main; the branch under test touched neither tick_triage nor its tests; the file passes green in isolation (127 passed, 50s).

Why this evades the Step 9c compare. The step-1d pristine single-file and paired-prefix oracle runs execute the file within minutes of collection, so the drift cannot reproduce there — the failures classify NEW (fail-closed) against an innocent branch, costing a manual provenance override (#2158 epm:test-verdict v1 records the full evidence chain). Every future long gate session (fleet-loaded VMs routinely stretch the ~18-min gate to 90+ min) re-rolls this die for whatever branch happens to be under review.

Fix candidate. Replace the module-level NOW with a per-test clock anchor: a function-scoped fixture (now = time.time() inside _event's caller, or _event computing time.time() - age_s at call time), OR monkeypatch/freeze the clock tick_triage reads so fixture ages are deterministic regardless of session wall position. Sweep the file for OTHER module-level time captures and any sibling test files using the same NOW = time.time() module-level pattern over age-sensitive predicates (repo-wide grep ^NOW = time.time() in tests/).

Acceptance. The 4 named tests pass when executed with a simulated ≥30-min collection→execution drift (e.g. monkeypatched clock), AND still pass in isolation; a regression pin prevents reintroduction of module-level wall-clock capture in this file.

Provenance

Filed by the /issue 2158 autonomous session after the Step 9c gate's compare classified the 4 drift failures NEW against an innocent branch (manual provenance override recorded on #2158, epm:test-verdict v1, 2026-08-18).