## What was red
`self-healing-advanced-triage`, `self-healing-agent-link-drift`, and
`self-healing-starved-refinement` — **7 failed / 19 passed** on clean
`origin/main`. All three fail the same way: the sweep returns `0`
recoveries where the test expects `1`.
## Cause
Those three sweeps were converted from `listTasks({ column: "triage" })`
to **role** filters (`filterByPreWipRole(..., ["intake"])`). Each store
fake has no workflow-selection readers, so the sweep resolves the
**default IR** — in which `triage` is not a declared column at all. The
seeded cards carried no intake role, the filters returned no candidates,
and the sweeps did nothing.
The product change was correct; the fixtures were asserting against a
column id the product had stopped declaring. `self-healing.ts` even
warns about this exact shape in its own comment: *"Converting a
predicate while leaving its source query on a literal produces a sweep
that LOOKS converted and does nothing."* The fixtures are the mirror
image — a seed on the old literal makes a correctly-converted sweep look
broken.
**Verified, not assumed.** I resolved the default IR and printed the
flags rather than reasoning from column names:
```
todo {"intake":true,"hold":true,"resetOnEntry":true}
in-progress {"countsTowardWip":true,"abortOnExit":true,"timing":true}
in-review {"mergeBlocker":true,"humanReview":true,...}
done {"complete":true}
archived {"archived":true,"hiddenFromBoard":true}
```
`todo` is the intake column post-U11 (the merged Planning column).
`triage` does not appear.
## Measured
| Check | Before | After |
|---|---|---|
| the three files | 7 failed / 19 passed | **26 passed** |
| all 40 `self-healing-*` files | 3 failed files / 7 failed tests | **40
passed / 695 tests** |
| `pnpm test:gate` | — | **726 passed** |
| `pnpm lint` | — | clean |
Restoring the deleted column id reintroduces the failures, so the seeds
are load-bearing rather than cosmetic.
**Pre-existing, untouched:** the same 9 vitest *"unhandled errors"* and
the identical non-zero exit appear on clean main. Verified by reverting
only these three test files and re-running — unrelated to the fixtures,
so flagged rather than folded in.
## One deliberately surgical spot
In `starved-refinement` only the starved refinement's own seed moved. My
first pass renamed the file's default seed and **broke a test that was
passing** — the auto-approve-all case calls `recoverApprovedTask`
directly (no role filter) and asserts a move **into** `todo`, so seeding
it in `todo` makes that move degenerate. I reverted and moved one seed
instead.
That leaves a real question I did **not** answer: post-U11 intake and
hold are one column, so a `triage → todo` move may no longer encode
anything. Deciding that means changing what the test is *about*, which
is a behaviour judgement on someone else's assertion. **Flagged in-file,
not guessed.**
Also worth noting for whoever converts the remaining backlog: line 43 of
that file pairs a `triage` seed against an explicit `todo` seed as a
contrast. U11 collapsed those two columns, so the contrast the fixture
was drawing no longer exists in the product — not a defect, but any
fixture built on intake-vs-hold being distinct is now suspect.
Census unchanged (722) — test files are not scanned by the census.
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
* **Tests**
* Updated self-healing workflow test scenarios to reflect the current
intake column.
* Improved test coverage for triage, agent-link drift, and starved
refinement handling.
* Added documentation clarifying workflow column and role-based
filtering assumptions.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
3d28e264c1