# workqueue-census — 2026-11-02

Typed census of every numbered work-queue (Q) item in `WORK-QUEUE.md`.

The census is a **snapshot**, and it says which one: `extract_manifest.json` pins
the queue revision and blob the rows were read from —

    source_rev       0ec1c643fa5cfcb996266423013de2b553eb0c95
    source_blob_sha  a7f716593b91c0db930d8ed7e78eb23a2d0f0d51   (WORK-QUEUE.md)
    source_sha256    ceb6cd3c7e0f1fbfe55045b5fd2765f076f7adae50f23b26f243e6e5028906eb
    source_lines     14265
    rows             88

(The README this replaces stated the source sha256 as `cea6cd3c…06eb`; the blob
pinned above hashes to `ceb6cd3c…06eb`. One hex character, in prose the gate did
not read — which is why the digest is now a graded leg rather than a caption.)

## Artifact

`workqueue-census.tsv` — tab-separated, header `Q | Title | Status | Target`,
one row per Q heading at the pinned revision. `Q` and `Title` are graded against
an independent re-derivation; `Status` and `Target` are transcribed from the
item's body and are **not** certified by this gate.

## Rules

* A Q item is a line matching `^## Q<n> — <title>`.
* Trailing `*(notes)*` on the title is stripped.
* `Status` is the first `**Status: ...` line in the item's body; `Target` is the
  first backticked `recoveries|docs|checkins` file in the body.
* Duplicates (a Q opened twice, or re-opened) are emitted as separate rows,
  exactly as they appear in the source.

## Reproduce / verify

```bash
bash CHECK                                 # grade the artifact against its pin
python3 extract_workqueue_census.py        # regenerate TSV + manifest at HEAD
python3 extract_workqueue_census.py --help # usage; a bad flag is rc 2
```

`CHECK` is a bash entry point over `grade_census.py` (the same shape as the two
check-in censuses that Q151 repaired): every caller in this tree invokes an entry
point with `bash`, and before this wrapper `bash CHECK` read rc 2 on shell syntax
errors without grading a cell. The grader is a separate program from the
extractor and does not import it.

`probe_census.sh` runs a 13-cell battery in a detached worktree, so the gate's
red readings are measured rather than asserted: a moved `Title` cell, a moved `Q`
cell, a corrupted digest, a pin whose revision does not hold that blob, an absent
pinned object, an absent TSV, an absent grader, a queue append the pin absorbs
(ADVISORY), the extractor refusing a dirty source, `--allow-dirty` writing an
unpinned manifest that the gate then refuses to grade, and a restore.

Exit codes: `0` green, `1` a fault, `2` could not grade (no pin, no artifact, no
grader). A `2` is never a pass.

## Column contract

* `Q` — the queue number (string, duplicate re-opens emit multiple rows).
* `Title` — heading text after `## Q<n> —`, trailing `*(notes)*` stripped.
* `Status` — the `**Status: ...` line, verbatim (may carry a trailing `**`).
* `Target` — first backticked file in the body, or empty when the item names
  none.

## What the pin does and does not buy

The queue is append-only, so a gate that re-derived from the live file reddened
on the next tick's own append — it graded the world, not the artifact (measured:
one appended heading turned the pre-repair `python3 CHECK` from rc 0 into rc 1
with three failing legs). Grading the pinned blob removes that time bomb: a
later append is reported as `ADVISORY source-moved` and changes no verdict.

The pin does not make the census current. Re-pinning is a deliberate
`python3 extract_workqueue_census.py` at a committed revision plus a commit; the
extractor refuses to pin a dirty `WORK-QUEUE.md` unless `--allow-dirty` is
passed, and that path records `source_blob_sha: null` — which the gate reads as
ungradable, not as green.