47. Editorial language scope, and the thirty rows their own sources already claim

Context

The request was to widen German and French support across the catalogue. The measurement that answers it says the gap is almost entirely not a capability gap, and that the two fields which look like one question are two.

core/describe.py already draws the line, and draws it deliberately:

• meta.languages is authored editorial scope — whether the form is a form in that language at all. wechselsatz is German by nature, not by capability.
• runs_in is computed from pack capabilities — whether this install can check the row in that language. Derived rather than authored, in its own words, "so it cannot drift into the false claim a [en] row made once French began working".

Measured over the 126 implemented rows on 2026-09-21, with every extra installed:

	en	de	fr
runs_in (capability)	126	124	106
languages (editorial)	122	46	14

174 (row, language) pairs run today and are not editorially claimed — fr 92, de 78, en 4. Capability coverage is close to complete; editorial scope is what lags, and the two were being read as one number. Closing all 174 would mean asserting, 174 times, that a form belongs to a language. That is exactly the unsourced claim ADR 0015 exists to refuse, and no measurement supports it.

Decision

Widen languages only where the row's own source field is the evidence, and leave the rest as an open editorial question rather than a backlog.

Thirty rows meet that test for French: their source names a French work, author or tradition, they run in French on this install, and they did not declare fr. lipogram is the clearest — its source is Georges Perec, La Disparition (1969), the most famous French lipogram written, and it declared [en, de]. Each of the thirty was verified individually: describe.runnable was asked for fr row by row rather than inferred from the tranche, and every one came back runnable with nothing missing.

Eighteen of the thirty already carried a French name and a French definition and needed nothing but the declaration. Twelve did not, and the strings were written before the declaration was made — eleven names and twelve definitions.

The criterion is "the source names a French attestation", not "the form is French in origin", and five rows sit on the looser half of that: palindrome (Sotades, revived by Oulipo), snowball (Ausonius, named by Oulipo), pangram (traditional, catalogued in the Atlas), pantoum (Malay, adapted into French in the 19th century) and assonance_constraint (the binding device of Old French and Spanish epic). Each is a form French practises under a French name, and each is declared on that reading. Stating the reading is the point: a later reader can disagree with it row by row instead of guessing what rule produced the set.

The same test over German-sourced rows returns zero. Four rows name a German-language original — buchstabwechsel, denckring, wechselsatz (Harsdörffer and Kuhlmann) and ideenwuerfeln (Jean Paul) — and all four already declare de. German needs no tranche of this kind. That is the disconfirming half of the measurement, and it is why this decision covers French alone rather than the two languages the request named.

One near-miss belongs in the record beside that zero. single_sentence is sourced to "the form of Gertrude Stein's and Thomas Bernhard's long prose", runs in German, and declares [en]. Its source names a German-language author but not a German original — the form is traditional and its other exemplar is English — so it fails this ADR's test. It is the row a stricter or looser reading of the criterion would move first, in either language.

A row may not declare a language it cannot speak

Adding a language to languages without a names and a definitions entry beside it costs nothing and pushes the row into Description.untranslated — the silent per-field English fallback that the 2026-09-04 sweep made visible precisely so it would stop being invisible.

That rule was already guarded at the data level, and this ADR found the guard rather than needing one: tests/test_catalogue_quality.py:: test_every_declared_language_carries_a_name_and_a_definition has asserted it over the whole catalogue since nine backlog rows arrived declaring a German they could not say a word of. So the tranche's twelve French definitions were not optional politeness; the suite would have refused the declaration without them.

tests/test_declared_language_strings.py adds the other half, at the surface a caller actually sees: for every registered procedure and every language it declares, describe(id, lang=...) must report neither name nor definition as untranslated. The two would disagree if _text or _fell_back stopped agreeing with the data — a row whose strings are all present but which describe still answers in English passes the YAML check and fails this one. Both encode the rule rather than today's counts, and the new one was proved red by giving clerihew a de it has no strings for.

Consequences

Thirty rows gain fr, and languages for French goes from 14 to 44. runs_in is untouched: no capability changed, no distribution was added, and the two numbers moving independently is the distinction working.

The declaration cost 24 French golden cases, which was not foreseen. tests/test_invariants.py::test_every_declared_language_has_a_golden_case requires a fixture in every declared language, so languages: [..., fr] is a promise that the row has been run in French and not merely thought about. Six of the thirty already had one; the other 24 were written and each was checked against its own checker — verdict and violation list both — before it was committed. Three are externally sourced (two traditional French palindromes and the standard French font-test pangram) and the rest are constructed, marked as such.

That cost is the decision's best feature rather than its overhead. An editorial claim backed only by a source: line is an argument; the same claim with a French text the checker accepts is evidence, and the suite now refuses the first without the second.

belle_absente stopped being the README's example of the distinction, and had to. The README said it was languages: [en] while running in all three — true when written, and true only because its languages was wrong: the row is Perec's. clerihew replaces it and cannot go the same way, because the form is English by nature and no French capability will ever make it French.

The remaining unclaimed pairs stay unclaimed — 144 of them, fr 62 and de 78 and en 4 — and are documented rather than queued. They are not a coverage gap and should not be counted as one. ISSUE-remaining-editorial-scope.md carries the text for the follow-up issue: what the remainder is made of, what evidence would close a row, and why a sweep is the wrong shape for it.

One row that meets this ADR's own test is not in the tranche. villanelle is sourced to "Traditional; fixed in its modern shape by Jean Passerat (1606)", runs in French, already carries a French name and a French definition, and declares [en]. It was not on the list this branch was given and adding it would have widened an approved scope silently, so it is named here and in the follow-up issue as the first candidate of the next tranche instead. That a hand-built list of thirty missed one is the argument for the follow-up issue being a criterion rather than a list.