oficina / memoria Public

🤖 fix: the third review — a tone bug fixed twice wrongly, and a spec

amendment made without reading the spec

Nineteen mutations, thirteen survived, three of them live defects. The
two worst are both this project's named failure, committed while
repairing an instance of it.

**The callout tone, third attempt.** tokens.css pairs every kind: the
class spelling takes `data-tone`, the attribute spelling takes
`rst-tone`, and the file header says so in a sentence. Attempt one wrote
`rst-callout="danger"` — a value on the kind attribute, which nothing
selects. Attempt two, the commit that described fixing it, wrote
`[rst-callout] data-tone=` — half of one pair and half of the other,
matching neither — and its three new assertions passed because the
string `data-tone="negative"` does occur in the stylesheet, inside the
class selector standing next to the one needed. Every flash in the
product still rendered untoned.

The lesson is not about tones: **grepping for a string is not reading a
selector.** The gate now derives its vocabulary from `[rst-…` selectors
only, checks each *value* against `name~="value"` rather than against
"something somewhere takes a value", and rejects `data-tone` on an
attribute-spelled element by name.

That found two more, both pre-existing and app-wide: `rst-cell-mut` is a
class-only utility, so every `<p rst-cell-mut>` selects nothing; and
`rst-btn` takes primary, ghost and danger, so every button spelled
"quiet compact" renders as the default. **Neither is fixed here** —
which variant a quiet button should be is a visual decision, not a
defect with one repair — both are listed by name in the gate, dated, and
flagged for Paul.

**The ASKED counter is reverted, and §8.1's clause moved instead.** The
last commit added a third integer, arguing that §8.1's "the window
contains at least one NOT_VISIBLE" could not fail. The observation was
right; the remedy was wrong twice. §13 already carried the correct fix —
*require each counted asker to have `VISIBLE = 0` and `NOT_VISIBLE ≥ 1`*
— and §8.1 was amended without opening it. Per asker, all three clauses
collapse into one condition, "asked in the window" needs no storage
because a refusal is itself the evidence, and the only per-person record
memoria keeps stays the size §3.2 says it is instead of quietly gaining
a "how often did they query" count.

The window-global version had a live consequence the fixtures missed: a
24-hour outage plus one routine refusal of one Docs-only member fires
the refusal breaker — the exact case the prose says is excluded.

**`PruneTallies` was never called.** It had a comment saying a retention
rule nothing enforces is a sentence in a document. That is what shipped.
It runs on every wake now.

**The breaker could be unwired and the suite stayed green** — deleting
the `recordForBreaker` call, and the `StartCounting` loop, both passed.
The predicate and the wiring were tested separately and the join not at
all, which is how a breaker ships reading "cannot determine" forever on
a working instance.

The rest: the counter write moves to `context.WithoutCancel` (AGENTS.md's
rule for additions to derived state) and the comment claiming a dropped
counter yields "cannot determine" is replaced with what actually happens,
because it does not; gate 8's placement half asserted against a slice the
test wrote itself and is replaced by one reading a real search's notices;
two predicate clauses gained the fixtures that pin them; the retention
test's stated reason was false (day 7 is the oldest baseline day, not day
8); and an unreachable "cannot determine" case was removed rather than
described.

Twelve mutations re-run against the fixes, all red, none surviving —
including the two that survived the first pass. `NotVisible >= 1` was one
of them: with no zero-rows written it is unreachable, the same shape as
the counter it replaced, so rather than delete a correct rule the row is
now inserted directly by a test. `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