Published content may be written in the reader's language; the repository's own artifacts may not

77. Published content may be written in the reader's language; the repository's own artifacts may not

Context

project.english-only v1 said, in one sentence and without qualification:

Code, comments, commits, documents and governance records are written in English, with no exceptions.

It is a blocking rule and it is wired to scripts/ci/english-only-check, which decides the part an alphabet can decide: a letter that occurs in Czech, Slovak or Polish and never in English, or a whole non-Latin script, in any tracked authored file. The site's content tree is in scope. There is no debt baseline in this tree, so one hit fails the gate.

That rule was written when the statement was exhaustively true of this repository: every artifact in it was a working artifact — something a contributor or an agent reads in order to change the code. For a working artifact the rule's reason is exactly right, and this decision does not weaken it. A comment in another language is a second rulebook for whoever cannot read it, and the repository is worked on by people and agents who think in a language that is not English, which is precisely why the gate exists.

Since then the repository grew a second kind of artifact that v1 never had to distinguish: the published site. site/content-src/ holds pages written for a reader who is not working on this repository at all. A page addressed to a Czech reader, in Czech, is not a second rulebook for a contributor; it is the product speaking to its audience.

The trigger was concrete. Two long-form articles were written from one evidence base, one in English and one in Czech — not a translation, two independently composed pieces. The English one could be published. The Czech one could not, and neither of the two escape hatches beside the check applied: name is for proper nouns, fixture is for a file that uses a foreign string as data, and the check's own header states that neither exempts a sentence. Declaring the article as a fixture would have been a misuse of an exemption whose documentation forbids exactly that, and --write-baseline would have laundered it into known debt. The honest options were to change the rule or not to publish.

Decision

The repository's own artifacts are written in English, with no exceptions. Published content addressed to a reader may be written in that reader's language, under three conditions, each of which is decided by the gate rather than by review.

1. Location. It lives under site/content-src/, the authored source of the published site, or is the generated projection of such a page under site/content/. Nowhere else in the repository acquires this permission — not docs/, not .ai/, not a comment, not a commit message.
2. Declaration. It is declared in scripts/ci/english-only-allow.txt with a third kind, translation <path> <language> <why>, carrying the language and a required reason, in the same shape as the existing two kinds.
3. Self-declaration, checked both ways. The page itself declares its language in its front matter ([extra] lang = "<language>"), and the check refuses a declaration whose language does not match the page's own. A one-sided declaration — an allow-list entry nothing on the page agrees with — is a finding, not a pass.

The rule moves to version 2 with the distinction written into its statement. The failure behaviour is unchanged for everything that is not published content.

Consequences

The gate stays strict where it was strict. Everything the v1 check refused, the v2 check still refuses: a Czech sentence in a .rs comment, a Slovak word in a rule, a commit message in another language, a Czech paragraph in docs/. The permission is confined to a tree whose purpose is publication, and confined further to files that declare themselves.

The exemption cannot be acquired by being added to a list. Because the page must declare its own language and the check compares the two, an entry added to the allow list to silence a finding fails unless the page really is what the entry says it is. This is the same property the fixture kind gets from its required reason, made decidable rather than reviewable.

The site must stop asserting one language for every page. site/templates/base.html hardcoded <html lang="en"> on every page of the site. A published Czech page under that template would misreport its language to screen readers and to search engines, which is a real accessibility defect (WCAG 3.1.1), so lang now derives from the page's declared language and defaults to English. Reciprocal hreflang links between a page and its counterpart in the other language are emitted from the same front matter.

The rule's population grows by a category rather than by exceptions. The alternative considered was to keep v1 and publish only English. That was rejected because it makes the site's reach a function of an internal working-language rule, which is the wrong axis: the reason English is mandatory inside the repository — that a contributor must be able to read every artifact that binds them — says nothing about an article addressed to someone who will never open the repository at all.

What remains undecidable is unchanged and still stated. The check is an alphabet test. A Czech sentence spelled in ASCII passes it, in v2 exactly as in v1, and the rule stays wider than the gate. What v2 adds is that the permitted case is now typed and cross-checked instead of being impossible.