`taskUsesProvider` resolves a task's **active lane** to decide which
providers it is running on. The **planner** half was converted to traits
— its own note in that function describes this exact failure and says it
was fixed — and the **executor** and **merger** halves were left as
`task.column === "in-progress"` / `=== "in-review"`.
So on a renamed board an actively-executing card resolved **no
providers**, and a provider rate limit never paused it: the engine kept
sending work to the limited provider with that card.
## Measured
Renamed board (`building` = wip, `checking` = review), limit triggered
by a peer:
| lane | before | after |
|---|---|---|
| executor | `["FN-TRIGGER"]` | `["FN-TRIGGER","FN-PEER"]` |
| merger | `["FN-TRIGGER"]` | `["FN-TRIGGER","FN-PEER"]` |
The trigger was paused either way — but only through the
*always-include-the-trigger* fallback, and **that is what made the hole
quiet**: one task always gets paused, so the behaviour looks like it
works.
Resolved from the **same per-workflow IR cache** the planner lane
already uses, so this adds no reads. Both halves fail soft to the legacy
literal, so an unresolvable workflow behaves exactly as before.
## How it was found — the part worth keeping
A scan for *"legacy literal within a few lines of a role-resolved call"*
flagged this file. That is the same heuristic that produced #2670, and
it is now 2 for 3.
The first thing I suspected here was the `done`/`archived` terminal
filter. I wrote that fix, and **its revert stayed green.** That is not a
dead end — chasing *why* it would not go red showed the lane check
already excludes finished cards, so the terminal literal there is
genuinely redundant, and the real defect was one line over in the lane
check itself. **A revert that stays green is information: either the
guard is vacuous, or you are looking at the wrong line.**
I dropped the unprovable change and kept the provable one. The
`done`/`archived` filter is deliberately unchanged, with that reason
recorded in the test.
## Revert proofs, isolated
- wip lane back to the literal → **1 of 55 fails** (the executing peer)
- review lane back to the literal → **1 of 55 fails** (the merger peer)
Three paired cases pass under both, so "pause everything with that
provider" cannot pass for "resolve the lane": a card parked in the
renamed planning lane is **not** paused on an executor limit, a wip card
is **not** paused on a merger limit, and the legacy vocabulary still
pauses peers when no workflow resolves.
## Verification
- 55/55 `usage-limit-detector`; `pnpm test:gate` **71/71**; engine
typecheck clean; `pnpm lint` clean
- census unchanged at 748 — the fail-soft literals remain by design,
which is why the count is not the measure of this fix
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
da77e61118