oficina / memoria Public

🤖 fix: the review's thirteen findings, two of them live bugs

Fable 5.1 reviewed slice 1 with the mandate §1.7 requires — pick the
mutations the author did not. Twenty-two compiled mutations, **eleven
survived**, and two of those were not weak gates but shipped bugs.

**S3 was fired at the wrong route.** `Content` built
`/suite/s3?item_id=X&revision=N`; the module declares the route as
`/suite/s3/{item_id}?revision=` on the line that names the constant, and
its conformance harness — which every sibling is held to — fires exactly
that. In production every fetch would have 404'd, memoria reads a 404 as
"not team-visible, or gone", and §8.1's S3 breaker would then have
reported a *permissions* anomaly for a routing bug: the plausible empty
list on the one call that fills the corpus.

**And the test pinned the wrong shape too**, which is why it was green.
That is the finding under the finding: a fixture written from the same
model as the code cannot catch a wrong assumption, it records one. What
found it was somebody who opened the module.

**S3 truncated an oversized body in silence.** S1 and S2 get this free —
a truncated JSON does not parse — but a body is plaintext, so a 2 MiB
prefix of a larger document indexed happily as the whole thing with its
tail unfindable. It now reads one byte past the cap and errors.

The rest, each with its gate:

- **A sixth `Disposition` was born silent.** `Notices()` enumerated the
  loud values, so a new constant neither rendered nor noticed — and the
  AHEAD tripwire instructs the next person to add exactly that constant.
  Written as its two exceptions, a new disposition is loud until somebody
  deliberately quiets it. `numDispositions` lets the test iterate the set
  instead of a list of the five its author remembered.
- **Gate 23 was looser than the spec** and stayed green with `Deadline`
  widened to 1200ms and 1400ms. Now the spec's own fixtures: 3×300ms
  under 810ms, and 5s/10ms/10ms under 1.2s.
- **Two revisions of one item were never fixtured**, so keying the dedupe
  on the id alone survived. That mutation renders a withdrawn revision on
  a later one's permission — retraction-by-edit, undone.
- **"Several concurrent batches" had no timing at all.** Sequential
  batches within one sibling passed everything.
- **`Rule` no longer returns notices.** Its only caller recomputed them
  in its own order and discarded them, so the Outcome sat unused on a
  concurrent path — and moving the recording into the goroutines is an
  ordinary refactor that left every test green and is a concurrent map
  write. Deleting the return removes where that bug can live.
- **`Withheld` is gone.** It rendered the same notices in different
  words, so the page's copy depended on which of two renderers a caller
  reached.
- **`NewSiblings` refuses a duplicate or empty name**, since the name is
  the key a document's source resolves through.
- Notice counts pinned to passages, the cursor fixture given reserved
  characters, S3's 5xx path tested, the header check made a closed set,
  and four comments corrected that cited sections not saying what they
  claimed.

Twelve mutations re-run against the fixes: ten red immediately. The two
survivors were real: the escaping fixture could not tell escaped from raw
(a space alone is not enough — it wanted "?" and "#"), and the sixth-
disposition case needs two edits at once, which no single-edit mutation
sees. Both closed, plus a control proving a new constant is loud.

`make ci` green; `-race -count=2` clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H3VKtXnq6fSmM2F11Bjcsy