From my own #2839, re-measured today. Of that issue's SQL-literal sites,
this is the one that decides what a **live view** shows.
## The defect
`listTasksModifiedSinceImpl` backs the SSE watcher and modified-since
polling — the incremental feed the dashboard applies to its task list.
Its `includeArchived: false` branch excluded the literal `archived`:
```ts
conditions.push(sql`${schema.project.tasks.column} != 'archived'`);
```
On a board whose archive lane is named anything else, that predicate
matches **every** row and excludes nothing. Archived cards arrive in the
live feed and reappear on the board.
Nothing errors, and a full refetch filters archived rows by another path
— so the symptom is archived work that comes back until the next reload.
That gets reported as *"the board is flaky"*, not as a bug.
## The fix
`resolveProjectColumnsForRoles` seeds the legacy ids before adding
resolved ones, so the set is never empty and an **unconverted board
excludes exactly `archived` as before**. The literal stays as the
resolution-failure fallback, where excluding nothing would be worse than
excluding the legacy id.
## Surface enumeration — three of four sites are dead
Four sites share this invariant. Verified rather than assumed:
| site | status |
|---|---|
| `reads.ts:558` (SSE / modified-since) | **live** — converted here |
| `liveParentFilter` | **no references anywhere** in `packages/` or
`plugins/` |
| `listLiveTaskDocuments`, `listLiveArtifacts` | referenced **only** by
`taskstore-remaining.test.ts` |
That's why this PR converts one site rather than four — the other three
are production-dead, and converting dead code would add risk for no
behaviour.
## Measured
| check | result |
|---|---|
| new PG suite | **4 cases** — legacy control, the renamed defect, a
live-lane negative, and the forensic `includeArchived: true` read |
| mutation (force the legacy fallback) | fails **exactly** the renamed
case; the other three hold |
| six gates + `tsc` | green |
The negative case is the one that matters most: resolving the archive
role must not start excluding **live** work, or the board silently stops
updating for real tasks — a worse failure than the leak this fixes.
## One process note
I corrupted this file mid-session by mutation-testing it while
uncommitted: a failed restore left a half-applied block, and a later
`git checkout --` discarded the fix entirely. Both were caught by
re-grepping for the symbol rather than trusting the restore. The
reliable pattern is **commit first, then mutate, then `git checkout` to
restore** — which is how the proof above was actually run.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
511f5b7e2b