## Claim
`packages/engine/src/scheduler.ts` — the largest **unclaimed** cluster
(28). Triaged, **not converted**, for the reason below. Census
unchanged: **722 → 722**. No baseline movement is claimed, because
nothing was converted.
## Why not converted
Three of us independently hit the same wall on our first file — #2683
and #2684 (`self-healing.ts`), #2685 (helper coverage). This measures
the whole backlog **once** so the remaining workers don't each pay that
cost.
Two constraints gate conversion. Neither is visible in the per-file
counts the work order sorts on.
### 1. Location — the helpers aren't importable from 80% of the backlog
`isIntakeColumnRole` / `isPreImplementationColumnRole` /
`isHoldColumnRole` live in
`packages/dashboard/app/utils/columnRoles.ts`, a dashboard-**app**
module.
| Location | Guards | Share | Importable? |
|---|---:|---:|---|
| `packages/engine/**` | 316 | 43% | no |
| `packages/dashboard/app/**` | 150 | 20% | **yes** |
| `packages/core/**` | 148 | 20% | no |
| `packages/dashboard/src/**` | 78 | 10% | no |
| `packages/cli/**` | 24 | 3% | no |
| plugins | 6 | 1% | no |
**150 of 722 (20%)** can call them at all. Widening the helper *set*
(#2685, correctly) does not move this number — it's the module's
location, not its coverage. Core already exports `resolveColumnFlags`,
so a core-side predicate module would be *the same* abstraction made
reachable, not a new one. It is a prerequisite for the other 80% and
**not sufficient** — see below.
### 2. Flag scope — the binding constraint
A role predicate needs resolved trait flags. Most guards run in
functions handed a bare task row with no IR to resolve from. Threading
one in changes a signature and its call graph: a **behavior change, out
of scope**.
File-level proxy over the 572 non-dashboard guards: **339** in files
that reference an IR/flags resolver, **233** in files with none.
**That proxy overstates convertibility, and the overstatement is the
finding.** Reachability varies *within* one file, so a file-level
verdict is unusable. In my claimed cluster:
| Site | Context | Convertible? |
|---|---|---|
| `scheduler.ts:1690` | `resolveWorkflowIrById(...)` +
`resolveColumnFlags(c)` in the same block | **yes** |
| `scheduler.ts:231` | `isLegacyDependencySatisfied(dep: Task \|
undefined)` | no — task only |
| `scheduler.ts:341` | `shouldHoldActiveFileScopeLease(...)` | no — task
only |
So "convert the file" is not a unit of work that exists in this backlog,
and the rule *"the baseline must shrink by exactly your converted
count"* cannot be satisfied per-file until the count is per-site.
## A guard that must be skipped, not guessed
`packages/core/src/task-merge.ts:254`:
```ts
if (!options.skipColumnIdentityCheck && task.column !== "in-review") {
```
The parameter is `Pick<Task, "column" | "paused" | ...>` — no IR,
deliberately. The in-source FNXC comment records that callers who *have*
resolved the `merge-blocker` trait pass `skipColumnIdentityCheck` rather
than spoofing `{ ...task, column: "in-review" }`. The trait-aware path
already exists *beside* this literal.
Converting it wouldn't remove a legacy id — it would delete the fallback
the option was introduced to make explicit. **Flagged and skipped.**
## Suggested census upgrade (not done here)
Emit per-site whether trait flags are resolvable in the enclosing scope.
That turns the work order from "largest file" into "largest
**convertible** cluster" and makes baseline shrinkage predictable. I did
not touch `scripts/lifecycle-column-census.mjs` — it is the shared
instrument and changing it unannounced would invalidate everyone's
in-flight before/after numbers.
## Method correction worth propagating to every fleet worker
Claim-collision scans must compare a branch to its **merge base**, not
to `origin/main`. `git diff origin/main origin/<branch> -- <file>`
reports a difference when the branch is merely *stale* (the file didn't
exist at its base) — it flagged dashboard and CLI PRs as touching engine
E2E files. I nearly skipped a free cluster on that false signal.
`feature/code-organization-wave17` is excluded from collision checks:
**1556 files, 150 commits behind main, already `DIRTY`**. It must rebase
wholesale regardless, and counting it as a claim marks *every* cluster
in the backlog as taken.
Docs-only — no source, no test, no census change.
ae23be79f7