[feraremag2680]

What this is

The interface is translated into ten languages, but a good part of what a person actually reads in the Accessibility panel, the PDF/A list, the signature details, the print dialog and the edit history was still English in every one of them. Those sentences are worded by the layers below the UI, and notify_error says why they were left alone:

Engine and OS messages are not translated: matching their English wording would break silently whenever it changes.

That is exactly right, so this doesn't try to match the wording from the outside. It has each layer hand its own wording over, apart from the document's own words, and lets the UI look it up.

The arrangement

Each layer that words sentences gains three things, in its own vocabulary:

• pub const WORDINGS: &[&str] — every sentence it can say, written once, with holes in it ({n}, {p}, {}).
• said(&self, words: Words) -> String — says one of them: look the wording up through the closure, then fill the holes with the document's values in a single pass, so a file name containing {n} goes in as written.
• pub fn in_english(wording: &str) -> String — the closure that changes nothing, for the CLI, the automation layer and the tests.

The UI passes &crate::i18n::words, which is a catalog lookup. A wording with no catalog entry comes back as it stands — which is the English — so a catalog that is missing one shows an English sentence rather than a gap, and the nine catalogs this PR doesn't touch keep reading exactly as they do today.

Because the sentence is now data the layer publishes rather than a literal duplicated in the UI, it cannot drift: change the wording and the catalog entry stops matching, which is what the tests check.

What changed

Engine side. a11y::Finding, preflight::Issue, sign::pdf::Detail and print::PrintError carried pre-formatted English Strings; they now carry the document's values and name their wording. Edit::label() likewise returns a Label { wording, value }, listed in EDIT_LABELS / BATCH_LABELS. The lists are a11y::{WORDINGS, REPORT_WORDS}, preflight::WORDINGS, sign::pdf::{WORDINGS, kind::ALL}, print::WORDINGS and the engine's two label lists.

a11y::report_html takes a language tag and a words closure, so a saved report reads the way the panel read, down to its lang attribute.

print::PrintError now has a variant per thing a person can be wrong about (NoPages, OutOfRange, NotAPage), with the numbers apart from the words; select_listed used to format that same sentence into Invalid. Faults rather than mistakes — a file that cannot be read, a poisoned lock — keep their own text, and said passes them through.

UI side. a11y_ui.rs (findings and the saved report), standards_ui.rs, sign_ui.rs and print_ui.rs call said(&crate::i18n::words). No new tl! literal is introduced anywhere, so {simplified_chinese,french,russian}_covers_ui_literals are unaffected (a closure argument is not a literal).

Spanish. 107 entries added to es.tsv, covering all of the lists above, with its SHA-256 updated in ATTRIBUTION.toml.

Tests

• spanish_covers_what_the_engine_says — one catalog entry per wording in all eight lists, so the catalog cannot fall behind the engine without a test saying so.
• spanish_keeps_the_holes_in_what_the_engine_says — each translation keeps the same holes as the English, named ones included.
• print::tests::wordings_are_said — said(&in_english) == to_string() for every mistake, one wording per mistake, a fault's text is never translated, and a value containing {count} is not re-read as a hole.
• cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo xtask layers, cargo xtask assets, cargo xtask parity all pass.

Two unrelated, pre-existing failures show up on my machine (Windows, system language Spanish), both of them also on an unmodified checkout:

• i18n::tests::windows_display_language_query_returns_a_locale_tag — windows_ui_language() returns the preferred-language list joined with \n, so the assertion fails on any Windows with two or more display languages. Described in sign: read the .p12 files Windows writes (BER, not DER) #453.
• tests/accessibility.rs (both tests) — the kittest harness leaves the language on Auto, so on a machine whose system language has a catalog the shell is drawn translated and get_by_label("Accessibility Checker Options") finds nothing. LC_ALL=en_US.UTF-8 cargo test -p pdfcraft-ui-egui passes everything. Setting the language explicitly in the harness would fix it; I left it alone to keep this PR to one thing, and I'm happy to send it separately.

With the language pinned to English, cargo test --workspace is green apart from the first item above.

Adding another language

Translate the entries; nothing else is needed. docs/localization.md has a new section, Sentences the engine words, describing the arrangement, and the per-language gate is three lines if you want one.

🤖 Generated with Claude Code

[feraremag2680]

Noticed while merging main into this branch, and left alone so this PR stays about the sentences: docs/localization.md on main (line numbers below are main's) carries two paragraphs twice, each in an older and a newer form, which looks like a merge artifact rather than intent.

• Lines 15 and 16 both begin "Dialogs, panels, menus, notices and history labels go through the catalog (tl!)". Line 15 lists Italian and Arabic among the translations and says Czech and Brazilian Portuguese cover the menus; line 16 is the earlier version, with Brazilian Portuguese still among the translations and only Czech covering the menus.
• Lines 20 and 23 both begin "Japanese interface text uses BIZ UDPGothic from [craft-fonts]". The second one's list of languages that need only the bundled Latin faces is missing Italian.

Happy to send a separate PR deleting the two stale copies if that is useful.

[echelon]

Thank you, @feraremag2680! This has landed on main: "i18n: let the layers below hand their sentences over, so they can be translated". I merged main: the structure-tree rules had become a streaming visitor (#926), so your wordings were ported onto it, and the signature verifier now uses your kind::* constants with main's newer, stricter classification (#1099). Only Spanish translates the engine's sentences so far; other catalogs fall back to English.

Please join us on Discord if you haven't already: https://discord.gg/artcraft

[feraremag2680]

Sent as #1778, for the catalog paragraph only. The font paragraph below it is duplicated three times rather than twice, and #1484 already folds those into one, so I left them to it.