## How this was found — by re-testing a claim of mine
The learnings doc records "named legacy-id collections" as **measured
and clean**: 48 declarations, all fallback vocabularies, builtin column
lists, or already-converted seams. #3014 disproved that conclusion —
`TIME_INDICATOR_COLUMNS` was in that population and was a live defect.
So I re-measured over the shape that actually matters: **collections
used as a membership gate against a column.** Nine exist.
| site | verdict |
|---|---|
| `columnRoles.ts` ×2, `useSessionFiles.ts` | the no-flags fallback
*inside* the role helpers — correct by design |
| `branch-group-ops.ts` | seeds the legacy pair then unions resolved
lanes — already converted |
| `DocumentsView.tsx` | marked `DELIBERATE-LITERAL` fallback chain |
| `TaskCard.tsx` ×2 | fixed in #3014 |
| `plugins/…/reconciler.ts` | plugin with no trait source — same class
as #3003 |
| **`useTasks.ts`** | **no flags path anywhere in the file** |
## The defect
`useTasks` stamps `recentAgentActivityAt` only for cards in `{triage,
todo}`. The note at that set argues over-stamping is harmless because
every consumer re-checks for an intake lane before showing anything.
That's true, and it **only protects against false positives**. On a
board whose intake and hold lanes are renamed, the pair matches nothing
— so no stamp is ever written, and a correct downstream role check has
nothing to filter. The planning border and pulsing badge never appear
while the planner is actively working the card.
## The supplier ships with the seam
An optional resolver with no caller is the first failure shape in the
learnings doc, and my own gate would flag it — so `App` supplies it in
the same commit. `useBoardWorkflows` moved above `useTasks` to make that
expressible; it depends on `projectId` alone, nothing about tasks, so
reading it first is safe.
Remote rows deliberately get **no** flags — they belong to another
store, and local board-workflow metadata must never be applied to their
ids. That's the rule the footer index already follows.
## Measured
| check | result |
|---|---|
| `useTasks` suite | 124 → **126**, all green |
| reverting the gate to the legacy pair | fails exactly the renamed
case; the negative (renamed WIP is not planning) still passes |
| `App.test` + `useTasks` together | **269 green** |
| gates | all five green; lint and `tsc` clean |
## One observation I could not reproduce
The `App`+`useTasks` pair failed once, on a single unnamed test, and
passed on **four** subsequent runs including three consecutive. The
captured output showed jsdom URL-parse noise from `MissionManager`
fetches rather than an assertion failure, and the same pair is green on
unmodified `main`.
I'm not quarantining another file's test on one unreproducible
observation, but recording it rather than letting a green rerun bury it
— if it resurfaces in CI, this is the prior sighting.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
5bdb8a1102