**Test-only.** Three files, two commits. No production changes.
| File | Before | After |
|---|---|---|
| `self-healing-db-corruption` | 5 failed / 1 passed | **6 passed** |
| `self-healing` | 1 failed / 411 passed | **412 passed** |
| `worktree-pool` | 2 failed / 57 passed | **59 passed** |
All three are the same underlying story in different costumes: **the
assertion is watching a channel production stopped using**, or a step
that aborts before it can log at all.
## Commit 1 — the fake store was missing the health refreshers
`surfaceDbCorruption` *refreshes* health before reading the snapshot
(`FNXC:IncompletePgPorts 2026-07-26-20:45`, so PG connectivity is
re-checked instead of trusting an always-healthy sentinel). The fake
carried **neither** refresher, so the async branch fell through to
`this.store.refreshDatabaseHealth()` — undefined — and the step threw
before reaching dispatch. **Every assertion in the file was measuring
zero calls against a step that had already aborted.**
Both stubs are **no-ops on purpose.** Production ignores the refresh
return and reads `getDatabaseHealth()` immediately after, so the
snapshot mock stays the single source of truth. My first attempt
delegated them to `getDatabaseHealth`, which consumed a *second* value
per pass from the test that queues three `mockReturnValueOnce` snapshots
(one per `runMaintenance`) and broke its corruption → clear → corruption
ordering. Faithful beats convenient.
## Commit 2 — two more debug-level assertions
- **`self-healing`**: `"auto-archive: archived …"` is emitted at DEBUG
(`self-healing.ts:2747`); the test asserted `.log`. The mock already had
`debug` (from #2573), so only the target was stale.
- **`worktree-pool`**: both checkout-failure cases assert on
`console.error`, which is *correct* — `createLogger`'s `debug` writes
there. But debug is **gated on `FUSION_DEBUG`** (`logger.ts:43`), unset
under vitest, so the line was never emitted. One test is literally named
*"logs checkout -- failure at debug level"* while asserting a channel
debug could not reach.
Fixed by enabling `FUSION_DEBUG="worktree-pool"` for the suite and
deleting it in `afterEach` so the flag can't leak into sibling files.
**Deliberately not** fixed by re-pointing the assertions at another
channel — that describes whatever the code happens to do rather than the
behavior the test names.
## Verified each actually guards
A test that merely stops failing can still assert nothing, so every fix
was mutation-checked:
| Mutation | NEW failures |
|---|---|
| `surfaceDbCorruption` returns early | **5** |
| remove the auto-archive debug line | **1** — that test, only it |
| remove the checkout-failure debug line | **2** — both cases, only them
|
## Known residual, stated rather than hidden
`self-healing-db-corruption` **still exits non-zero** with 9 unhandled
`this.store.listTasks is not a function` rejections from
`openSurfacingCycle` (`self-healing.ts:7737`). These **predate this
change** — identical count before and after. The maintenance pass opens
one shared surfacing cycle up front, independent of which steps
`stubMaintenance` stubs.
I tried to clear them and backed it out, twice:
- adding `listTasks: async () => []` lets the cycle open, but then
*other* unstubbed sweeps run for real — an orphaned-planning-segment
audit fires and breaks 3 assertions expecting `recordRunAuditEvent`
never to be called;
- stubbing the four `surface-*` siblings didn't help either, because the
cycle is opened by the **pass**, not by the steps.
Making that file honestly green needs a fake complete enough for the
whole maintenance registry — a bigger change than the bug in front of
me, and one that would bury the fix above. Flagging it rather than
shipping a half-sweep.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
fb7ab6df26