Fleet phase. Claimed **`packages/engine/src/self-healing.ts`** — the
largest cluster at **56 of 126** total sites. Verified unclaimed first:
no open PR touches the file and no active worktree held a branch on it.
## Census before / after
| | total | self-healing.ts |
|---|---|---|
| before | **126** | **56** |
| after | **121** | **51** |
`census --strict` exits 0; baseline re-recorded in this commit so the
retired allowances cannot be regrown into.
## What converted, and why each role
`archiveStaleDoneTasks` asked "has this card finished?" by comparing
column ids, so on a renamed board it treated every finished card as live
and archived nothing — the sweep was inert on exactly the boards this
program exists to support.
- **active-dependents scan** and **temp-worktree age gate** →
`TERMINAL_ROLES` (complete ∪ archived): both ask "is this card done
with, in any sense?"
- **staleness filter** → `complete` **alone**: this sweep *archives*
finished cards, so an already-archived card is not a candidate. Using
the terminal pair here would have made the sweep consider its own
output.
**Union, not per-task, deliberately.** Over-inclusion is free at these
sites because the per-card check still discards, and the union needs no
per-task workflow selection — the failure mode
`resolveWorkflowIrForTask` has, where a card with no recorded selection
silently resolves to the built-in board. Recorded in
`docs/solutions/workflow-learnings/project-union-versus-per-task-lanes.md`.
## The half-converted state is the interesting part
Converting only the first two guards made `archiveStaleDoneTasks`
**register as a converted sweep** — the existing ratchet suite grew from
**36 to 38 tests** — and it then failed for still carrying `t.column !==
"done"`.
That is the failure mode worth naming: a partial conversion is worse
than none, because the function now *looks* converted (it calls the
resolver, it reads as role-aware) while one guard still pins it to the
legacy vocabulary. Finishing the function turned it green. I would not
have caught it from the diff.
## Verification
- `self-healing` suites — **807 pass** (41 files)
- `tsc --noEmit` — **0 errors**
- `census --strict`, `check:lane-wiring`, `check:fnxc-future-dates`,
`check:inert-flag-seams`, `check:sql-column-literals` — all exit 0
## Flagged, not guessed — the remaining 51
Deliberately left, each for a stated reason rather than an omission:
1. **Move-transition matrices** (~1489–1504): `from`/`to` pairs encoding
a legal-transition graph (`in-progress → todo|in-review|done|archived`).
These are the *shape* of the lifecycle, not a lane lookup; converting
them needs a transition-role model that does not exist yet. Guessing
here would encode a wrong graph.
2. **`getLiveTaskColumn` comparisons** (~1398, 5313–5342): compared
against a normalizing accessor that manufactures `"archived"` for
soft-deleted rows. Those are protocol values, not column ids —
converting them changes what the sentinel means.
3. **Sites without store access** in scope (several module-level
predicates): need the resolved set threaded in as a parameter, which is
a seam change per call site, not a substitution.
4. **`todo` requeue targets** (1927, 6181, 6303, 12060–12064): these
pick a destination, so they want the single `intake`/`hold` answer from
`resolveLifecycleColumns`, not a set — different arity, and several are
inside sweeps whose rebound semantics I would be changing rather than
preserving.
Each is a real conversion; none is a one-line substitution, and doing
them blind is how a guard count drops while behaviour gets worse.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
c9516dbd09