## How this was found
Full-suite **shard 3/4** reports no `Tests N failed` summary at all —
its log ends mid-`@fusion/engine [1/2]` on a watchdog heartbeat, so the
red reads as infrastructure noise. It is not.
Facts that ruled out the infrastructure explanations, before touching
any test:
- watchdog budget is **1500s**; the engine slice died after **~6.5–10.5
min** (varies run to run) — not a timeout, and not a fixed one
- job `timeout-minutes: 60`, ran **9.7 min** — not the job timeout
- `concurrency.cancel-in-progress: false` — not cancellation
- annotation says **exit code 1**, not 137 — not an OOM kill
- across **four consecutive runs** the last file named is always
`reliability-interactions/explicit-duplicate-marker-sweep.test.ts`
Running that file locally reproduces real failures. The summary line is
simply missing from the CI log's final chunk (the last ~5s of output
never appears), which is what disguised a normal test failure as a
crash.
## Four causes, none a product defect
| File | Failures | Cause |
|---|---:|---|
| `explicit-duplicate-marker-sweep` | 2 | fixture seeds `column:
"triage"` |
| `starved-refinement-x-approval-gate` | 1 | same |
| `starved-refinement-x-triage-poll` | 1 | same |
| `executor-pending-review-skip-retry` | 1 | review handoff now passes
move **options** |
`triage` is no longer declared on any workflow post-U11, and these
sweeps filter by **role** — so cards seeded there carried no intake role
and the sweeps reported 0.
## Measured
| Check | Result |
|---|---|
| `reliability-interactions` | 5 failed → **0** (103 files, **530
passed**) |
| `pnpm test:gate` | **726 passed** |
| `pnpm lint`, engine `tsc --noEmit` | clean |
Census unchanged — test files only.
## The real find: "honors the disable flag" could not fail
Forcing `enabled = true` in `resolveExplicitDuplicateMarkerTasks` — i.e.
making the sweep **ignore the disable flag entirely** — left all 16
cases **green**.
The fixture never set `triageDuplicateResolution`, so the sweep had no
resolution action to take and the duplicate survived whether the flag
was honoured or ignored. The assertion held for a reason unrelated to
the test's name. It had been red only because of the column literal,
which would have made "fix the literal, go green" a repair that left a
guard guarding nothing.
Setting the resolution mode makes that same mutation delete the
duplicate and the case fail. Verified both directions:
| | mutation applied |
|---|---|
| before | 16 passed — **guard cannot fire** |
| after | **1 failed** / 15 passed |
Recorded in-file with the measurement, so nobody strips the setting back
out as redundant.
## Two assertions strengthened rather than relaxed
- The disable-flag and failed-delete cases now assert the column is
**unchanged from a value read before the sweep**, rather than equal to a
literal. They cannot pass because a seed happened to land where the
assertion looked, and they survive the next column rename. The
invariants those cases own are "the flag stops the sweep" and "the task
whose delete threw survives" — the column id was always incidental.
- The handoff move asserts its **provenance** (`nodeId:
"review-pending-handoff"`, `preserveProgress: true`) instead of the
`expect.anything()` its siblings in that file use, so a move to the same
column by another path cannot satisfy it.
## Still open for whoever owns CI
Shard 3/4's log loses its final chunk, which is why a plain test failure
presented as a crash and stayed unexplained across at least four runs.
Anyone triaging full-suite from shard conclusions alone will keep
mis-reading this one; the failure has to be reproduced locally to be
visible.
a037ca93c7