Translations: validate catalog structure, repair Tamil

[JumpLink]

Follow-up to #155, which added Tamil.

Why

That PR arrived at 456/456 translated and merged green, because 456/456 is what msgfmt measures: a msgstr counts as translated when it is non-empty. Two things it never looks at had come along with it.

Six Tamil strings are not well-formed markup. Tutorial paragraphs render through GtkLabel use-markup, so Pango treats those as parse errors and the reader loses the paragraph. Tamil was the only catalog with any. And ten entries carried translated code literals — thirteen <tt> spans, nine distinct: <tt>checkSnakeCollision</tt> became <tt>செக் பாம்பு மோதல்</tt>, <tt>0001 AND 0001</tt> turned the AND opcode into the Tamil word for "and", and <tt>JSR end</tt> became <tt>JSR முடிவு</tt>, which no longer assembles. The tutorial tells the reader to find these in the editor.

Checking the other fifteen catalogs found the same class already on main, in seven of them. So this PR fixes the class, not the one instance, and adds the check that would have caught all of it.

What changed

Seventeen entries across seven catalogs — de 1, fr 1, he 1, ja 6, uk 1, vi 2, zh_Hans 5. vi had translated JSR loop and BNE decrement; zh_Hans had swapped hyphens for en dashes inside hex ranges and dropped a sentence about the 16 colours; ja had lost four <i> spans and spinWheels. Surrounding translations left alone.

Tamil. Markup repaired, code literals restored, four half-English entries translated. The word "assembly" appeared as three different wrong words: பேரவை (8x) and சட்டசபை (3x), both a legislative assembly, and கூட்டி (an adder, 1x), next to the correct அசெம்பிளி (8x). It reached the application name and the file menu:

English	was	now
Learn 6502 Assembly	6502 சட்டசபையை அறிக	6502 அசெம்பிளியைக் கற்றுக்கொள்
Assembly	கூட்டி	அசெம்பிளி
Assembly Files	பேரவை கோப்புகள்	அசெம்பிளி கோப்புகள்
Open Assembly File	பேரவை கோப்பைத் திறக்கவும்	அசெம்பிளி கோப்பைத் திறக்கவும்
Save Assembly File	பேரவை கோப்பை சேமிக்கவும்	அசெம்பிளி கோப்பைச் சேமிக்கவும்
Step-by-step guide to 6502 assembly	6502 சட்டசபைக்கு படிப்படியான வழிகாட்டி	6502 அசெம்பிளிக்கான படிப்படியான வழிகாட்டி
About Learn 6502 Assembly	பற்றி அறிய 6502 பேரவை	6502 அசெம்பிளியைக் கற்றுக்கொள் பற்றி

Same pattern in the prose: "to" as an e-mail recipient (பெறுநர்), "lead" as the metal (ஈயம்), "current" as electric current, "logic" as a lawsuit, and "van der" as a delivery van inside a surname. 32 entries changed of 456.

The submission's house style is preserved. The catalog contains no grantha letters at all, writing s/j/z with ச (ச்டாக், ச்னேக், ச்பானிச்), and the new text follows that.

packages/translations/check.js, wired into gjsify workspace @learn6502/translations check and run in CI before the type checks. Four rules: markup parses, no tag the source opens is dropped, <tt> content survives byte-for-byte, format placeholders survive.

Tags are compared by name, so a translator can still point an <a> at a target-language article as the README allows. Only in-document anchors are pinned, since a rewritten #jumping is a dead link rather than a localisation. One <tt> literal is exempt with its reason written down: label in the addressing-mode list is a placeholder word, not an identifier.

Calibrated against all sixteen catalogs: on main it reports 32 strings across eight catalogs — the 17 above plus 15 in Tamil — and nothing else. Reverting any one of those 32 fixes makes it red again, on both the Node and the GJS path. The remaining Tamil edits are terminology, which no structural rule can see.

Worth flagging, because the first version of this check was green for the wrong reason. It located the catalogs relative to its own file, and under --app gjs it runs as a bundle in dist/, where there are no .po files. Every rule passed over an empty set and the run reported success having read nothing, while the same code under Node correctly went red. The catalog directory is now the working directory, and the check compares the number of catalogs it read against LINGUAS so that reading nothing is loud rather than green.

LINGUAS sorted (ta had been appended after zh_Hans), and the README gained the Tamil row plus the <tt>-is-code rule and the check command.

Review pass

The check was then attacked deliberately, and three findings came back.

It passed content GtkLabel renders as an empty string. Measured against GTK 4 / Pango 1.57, the first version accepted a bare &,  , a tag Pango does not know (added alongside the source tags, so the multiset comparison stayed happy), an attribute a tag does not support, <tt class="c"> — which also stops the <tt> literal from being compared at all — and an <a> with the href dropped. It also accepted a placeholder added by the translation, and rejected <a href='x'>, which GtkLabel accepts. Rule 1 is now a model of the vocabulary a label takes, and rule 4 compares placeholders in both directions.

The PO reader skipped msgstr[0] silently. No catalog has a plural yet, so nothing was wrong today; the day one is extracted, all four rules would have passed it without looking. msgctxt and plural forms are read now. The catalog guard compares names against LINGUAS rather than counting them, so a stray .po beside a missing one cannot keep the total right while leaving a language unchecked.

Nothing proved the rules still bite. 21 pairs the rules must classify correctly run before any catalog is read, and weakening the tag allowlist, the entity pattern, the attribute table, the <tt> comparison or the added-placeholder check each turns the run red on the self-test alone.

Three Tamil entries of the repaired class were still on the branch, found by measuring how much of each msgid survives verbatim in its msgstr — the structural check cannot see them, because the markup and the <tt> literals are intact and only the prose is untranslated:

Games now automatically இடைநிறுத்தம் when switching between screens
<b><tt>$10</tt>-<tt>$15</tt></b>: Snake position தகவல்கள் (head, body segments, tail)
இதோ ஒரு சான்று. குறிப்பு that immediate operands அரே still prefixed with a <tt>#</tt>.

The last one also carries அரே, a transliteration of the English "are". Every other catalog translates all three. Each replacement is built from words the catalog already uses and keeps the house style, but the wording needs the same validation as the rest of ta.

What this does not do

The check says nothing about language quality. Tamil is listed as Needs validation, like the other AI-assisted catalogs.

@TamilNeram, thank you for the Tamil translation. The machine pass that produced parts of it damaged some markup and mistranslated "assembly", and I have repaired what I could verify mechanically while keeping your spelling conventions. Would you be willing to check the wording? The table above is the place to start. Two open questions where I picked a form and would rather you decided:

• register is both பதிவு (19x) and பதிவேடு (9x) in the catalog, instruction both வழிமுறை (19x) and அறிவுறுத்தல் (12x), label both லேபிள் (8x) and சிட்டை (6x). I used the more frequent one in the entries I touched and left the rest alone.
• "most significant byte" I rendered as உயர் வரிசைப் பைட். If there is a better established Tamil term, please say so.

Verification

• gjsify workspace @learn6502/translations check passes on all sixteen catalogs
• msgfmt -c passes on all sixteen, 456/456 each, no fuzzy
• gjsify format --check and gjsify lint clean for the added file
• built and ran the GNOME app with LANGUAGE=ta

The second tile now reads 6502 அசெம்பிளிக்கான படிப்படியான வழிகாட்டி, where it read 6502 சட்டசபைக்கு … before.

I did not drive the app to the tutorial page, so the repaired tutorial markup is verified by the parser rather than by eye.

Noticed while testing, not part of this PR

With LANGUAGE=ta set and the main window in Tamil, the About dialog shows the German application name and English row labels. I have not looked into why, and it is unrelated to the catalog changes here. Flagging it in case it is worth its own issue.

🤖 Generated with Claude Code

https://claude.ai/code/session_01URLYj1wgGxVNRbBpNT8xAV