Taking `packages/core/src/live-agent-count.ts` from the shared
triage-guard backlog. **This PR does not convert it** — it corrects a
comment that would have stopped the conversion, and records why the
conversion is not a one-liner.
## Per-file guard count
| File | Before | After | Note |
|---|---:|---:|---|
| `packages/core/src/live-agent-count.ts` | 2 | **2** | not converted —
see below |
Census across `packages/*/src` excluding tests, for the pattern `column
=== "triage"` / `column !== "triage"`:
| File | Sites |
|---|---:|
| `engine/self-healing.ts` | 10 |
| `dashboard/src/routes/register-task-workflow-routes.ts` | 7 |
| `core/task-store/comments-ops.ts` | 3 |
| `dashboard/src/routes/board-workflows.ts` | 2 |
| `core/task-store/task-creation.ts` | 2 |
| `core/live-agent-count.ts` | 2 |
| `engine/spec-staleness.ts`, `engine/replan-target.ts`,
`engine/mission-feature-sync.ts`, `core/types/archive-planning.ts` | 1
each |
## The finding
The comment in this file asserted the literal fallback was unreachable:
> *"The literal fallback is fixture-only; board/store callers always
supply flags/IR."*
**It is false.** `useExecutorStats` resolves
`columnFlagsByTaskId?.get(task.id) ?? columnFlagsById?.get(task.column)`
— `undefined` for any card whose column is absent from the board's flag
map, which is exactly the renamed-or-undeclared column case. So the
literals run in production, on the cards least likely to match them.
**Consequence is under-reporting, not a stall.** A card in a renamed
planner column matches neither `triage` nor `todo`, so
`isWaitingAgentTask` returns false and the footer's queued count
silently omits it. Default-workflow cards still match through the `todo`
arm after the Planning merge, which is why nothing looks broken — the
same "still fires via the todo arm" shape as the executor sites I
audited in #2572, but here with a real observable effect.
## Why I did not convert it
Removing the id guesses means deciding what an **absent flag set**
should mean, and `"not intake"` is as much a guess as `"todo is intake"`
— either choice moves the numbers the operator sees in the footer. Doing
that safely needs the dashboard's flag-map population understood and a
test that pins the queued count, neither of which is a small change.
A comment asserting an untrue invariant is worse than no comment: it is
precisely what would stop the next person converting these two sites,
because they would read it and move on. Correcting it is the useful part
I can land with confidence right now; the census entry stays open.
## Verification
`tsc --noEmit` on `@fusion/core` clean; `pnpm lint` clean. Comment-only
change to production source, so no behaviour change and no changeset.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
71f64025d8