Claiming `packages/engine/src/executor.ts` from the census work order.
## Census before/after
| File | Before | After |
|---|---:|---:|
| `packages/engine/src/executor.ts` | 7 | **4** |
Measured with `scripts/lifecycle-column-census.mjs` (kind `column`
only), not grep.
## Converted (3)
**L17258 — the completed-task watchdog never armed on a renamed board.**
It required the card to sit in a literal `in-progress`. This does not
error; the watchdog simply never fires, which is the silent-guard class
this program exists to remove. The branch immediately above already
resolves the same lane through `resolveWipTargetForTask`, and there is
even an FNXC note there saying `latestColumn` must come from that
resolved value — so the comparison now asks the same resolver rather
than an id.
**L14940 (×2) — the duplicate-handoff finalize never ran on a renamed
review lane.** `fromColumn`/`toColumn` are parsed out of the store's
rejection message (`Invalid transition: 'X' → 'Y'`), so they carry
whatever ids that workflow declares. Comparing them to the literal
`in-review` meant a renamed lane never matched and
`finalizeAlreadyReviewedTask` was skipped, leaving the card
mid-transition with nothing to complete it. Now resolves the task's own
review role, falling back to the legacy literal when the workflow cannot
be read — so behaviour is unchanged wherever the vocabulary is
unreadable.
## Flagged, not converted (4) — per the fleet rule that behavior changes
are out of scope
**L3557 / L3581 / L3632 / L3642** are branch conditions inside the
**synchronous** `store.on("task:moved")` listener. Resolving a task's
workflow requires an `await`, which is not available in a sync
listener's condition. Moving the test into the deferred body would widen
the branch to every non-forward move and then re-narrow it — a
**behaviour change to the planning-evacuation path**, not a vocabulary
conversion. Converting them properly means making the listener async,
which wants its own commit and its own test.
I flagged rather than guessed, which is why this is 7 → 4 and not 7 → 0.
## On test coverage, stated plainly
Both converted sites are pure resolver swaps in `async` contexts,
verified by tsc, the census delta, and the existing executor suites (48
tests green). I did **not** add new fixtures: this is the file where I
twice wrote tests that passed against the *unconverted* code —
`recoverCompletedTask`'s seven early-return guards make negative
assertions succeed trivially — and reverted both times rather than claim
coverage I did not have. A fixture that genuinely drives L17258 needs a
satisfied `workflowStepResults` so the run does not divert into graph
re-entry; that is worth doing, and it is worth doing honestly rather
than as a green-looking placeholder.
## Verification
- census: `executor.ts` 7 → 4
- `tsc --noEmit` on `@fusion/engine` clean; `pnpm lint` clean
- `executor-graph-boundary`, `executor-task-done-summary`,
`executor-triage-column-audit`, `executor-step-session` — 48 tests green
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
eb0ee4ae98