A patch, and the number is the honest one. A minor closes an arc. Two arcs each gained their first chapter here — 1.59, a language needs a maintainer, and 1.60, our own medicine, measured — and neither is closed. Nothing here is a new command; both chapters are the same act done twice: a sentence this project had been saying about itself turned out to be wrong, and was measured rather than argued with.
Added#
Five of the seven phrase dictionaries were written by nobody who reads the language, and now the report says so. The first chapter of arc 1.59 — a language needs a maintainer. The dictionaries cover English, Spanish, French, German, Portuguese, Italian and Dutch, and the coverage line named all seven in one sentence, which reads as seven dictionaries of equal standing. Two of them are languages this project reports in, which is the only evidence in this repository that anybody here reads them. For the other five, nothing in the history says a speaker ever agreed that removing an entry leaves the prompt asking for the same thing.
The roadmap and the catalogue disagreed, and the catalogue is the one users meet. An eighth language has been held back for several arcs on the stated grounds that a dictionary is a judgement about a language and this project will not make it in a language nobody here reads. Seven dictionaries shipped anyway.
The evidence that reading the list is not enough is a bug this project already
shipped: INTENSIFIERS carried molto, muito and heel, each an
intensifier and a quantifier, so you have much time became you have time
in three languages at once. Spanish avoids exactly that trap — muy yes,
mucho no — because somebody who speaks Spanish wrote it. The three were caught
by running prompts through the rules, which is a far weaker instrument than a
speaker, and one bug found by the weaker instrument is not a review.
DICTIONARY_STANDING records reviewed or unreviewed per language with
what was actually done to the entries, and trazum optimize prints the
unreviewed set on the branch where an empty result would otherwise reassure —
its own line, in both report languages, so the day somebody maintains all seven
it is deleted rather than reworded.
docs/language-maintainer.md is the role made
real: what a maintainer decides (five questions, none of them "is this the right
translation"), what is asked of them, what is deliberately not asked, and what
happens when nobody holds it — the language stays, its record says unreviewed,
and nothing pretends otherwise.
It is also a row on our own medicine, and a different shape from the others there. Every other row is a claim nothing checked; this one is a rule this project wrote for itself and then broke, held up for several arcs as the reason an eighth language was not scheduled while five unread dictionaries were already shipping. No guard catches that — only re-reading the rule against the catalogue does.
What this chapter does not do is promise an eighth language. Whether the role is ever filled is not a scheduling question, and the page says so.
Nothing is deleted. A Dutch prompt is better served by a dictionary that fires and says it was never reviewed than by silence.
The guards are derived, not written: maintainers.test.js reads the reviewed
set off the report catalogues on disk, so a French report translation fails the
build until somebody decides what it means for the French dictionary; the
document is checked against the code's unreviewed set rather than against a
list; and both checks are proved against a fabricated table and a page with a
language quietly dropped.
Changed#
One of the three things this project could not say about itself has stopped being true. docs/our-own-medicine.md ends with three admissions, and arc 1.60 asks for at least one of them to fall to a measurement rather than an argument. The third was "every miss above was found and written down by the same process that made it".
Five defects on the record were found by CodeQL — in 1.8.0, 1.46.0, 1.50.3, 1.53.4 and 1.55.0 — and not one by a test in this repository. They are tabulated on the page with what each found: an SSRF where a validated URL and a fetched URL were different expressions, a time-of-check/time-of-use race on a size bound, two unanchored host patterns that a lookalike domain satisfies, a ReDoS in a guard this project had just written whose own proof would have passed against the vulnerable version, and a file-system race on the pull request that introduced it.
The 1.8.0 entry is the one that carries the weight: CodeQL kept that alert open twice, against this project's judgement, and was right both times. A self-report cannot contain that shape by definition.
And what it does not establish is written beside it. CodeQL is not an independent audit — it runs because this project turned it on and would stop the day somebody here deleted a workflow. It is an outside instrument whose rules this project did not write and cannot argue with, which is narrower than "somebody audited us" and is what was actually measured. The other two admissions — no usage log of its own, no outcome recorded for any of its work — are untouched, and a test asserts the page still says so.
The list of releases is guarded: docs.test.js fails if one named in the table
stops mentioning CodeQL in its notes, so the claim cannot grow past its evidence.
Fixed#
docs/releasing.md listed two of the three places @trazum/core is pinned.
The recipe named packages/cli and packages/mcp; apps/web pins it as well,
so anybody following the steps left the web app resolving a registry copy of the
previous version instead of the workspace. Found by publish.test.js while
cutting this release, and the recipe now says three.