A minor, and it closes an arc out of order. 1.57.0 and 1.58.0 are still open and stay open: 1.57's remaining chapter needs a provider credential this repository does not have, and 1.58 is an editor extension, which is a distribution commitment rather than a feature. An arc that can be finished is worth more than a slot left idle waiting for one — the same call 1.55.0 made, for the same reason, and the gaps stay gaps rather than being renumbered away.
The arc asked for one thing and got it: make the maintainer requirement a real, documented role with a real bar, and then admit that whether it lands is not a scheduling question. Three chapters did that. What it deliberately did not do is add an eighth language.
The uncomfortable half is that the arc's own premise had been broken since before the arc existed. The rule was a dictionary is a judgement about a language and this project will not make it in a language nobody here reads. Five such dictionaries were already shipping. That is now a row on our own medicine and a rule in the doctrine.
Added#
And the worklist that role asks for can now be printed. The third chapter of
arc 1.59. The page asks somebody to read one dictionary's entries and judge them
one at a time — and until now that request could not be scoped, because the
entries live in one flat array per rule with a // Dutch comment marking where
each language starts. A prospective maintainer had to read phrases.ts to find
out how much they were agreeing to.
node scripts/dictionary-worklist.mjs nl # or fr, de, pt, it
node scripts/dictionary-worklist.mjs fr --jsonThe grouping was real and existed only as a comment. phrases.test.js
already parsed those markers to check no language section was thin, so the
structure was derivable and simply unavailable to anybody outside this
repository's test suite.
Two things the counting found. The page said "a few hundred short phrases" and it is thirty to thirty-eight — an afternoon, not a project, and the figure was wrong in the direction that discourages volunteers. And English has 89 entries against Spanish's 81, so the five unreviewed dictionaries are also less than half the size of the two somebody read; that is recorded on the page and not addressed.
The guard is a second, different parse. The worklist slices between language markers; the check counts the whole array and fails if the two disagree — an entry written above the first marker belongs to no section, would appear on nobody's worklist, and would still be editing people's prompts. Proved against a fabricated dictionary with exactly that shape. The five counts the page quotes are checked against what the script produces, so the number a volunteer is shown cannot drift from the work.
Found by running it: piping the output into head, which is what the
documented usage invites, printed an EPIPE stack trace from a script that had
already done its job.
The warning now reaches the branch where it matters most. 1.56.2 said which five of the seven dictionaries nobody here reads, on the branch where no rule fired — and that is the branch where the prompt is untouched. The second chapter of arc 1.59 puts it where the change is: when a rule fires on a prompt whose own language is one of the five, the report says the entries that just edited it were never agreed by a speaker, and to read the diff.
Gated on the prompt's own detected language, so an English or Spanish prompt
never sees it and it does not become a footer under every report.
detectTextLanguage answers null on a prompt too short or too mixed to place,
and this stays silent then: not-detected is not not-unreviewed, but guessing a
language in order to warn about it would put a Dutch warning on a Portuguese
prompt — the same overreach the detector exists to refuse.
Proved by running, in both directions: a Dutch prompt that fires the filler entries gets the line, an equally wordy English prompt does not, and a Dutch prompt too short to place fires a rule and still gets nothing.
Changed#
A rule joined the doctrine, and it is the only one there that no test can enforce. A rule you wrote for yourself is a claim like any other — learned by breaking it. The rules in that document are enforced by tests because a rule with nothing checking it drifts as fast as a number with nothing checking it; the ones about the project's own conduct drift faster, because a test asserts what the code does and nothing asserts that the project still does what it said it would.
The cheapest available check is written down with it: put the promise next to the inventory. This one was found by writing the rule and the catalogue into the same document and noticing they did not match.