A language folder is drafted from the English master, and proved complete

[maximsan]

Why

The drafting half of the translation step — ADR-0026's three jobs: draft a
language folder, mark it complete, queue what has gone stale. Closes #603.

The words are agent-drafted, so the tool does the bookkeeping around them:
plan writes a queue of every piece of prose that is absent or stale, an agent
fills each text, apply writes the folder with a fingerprint per field, and
check refuses to call a language complete while anything is missing.

Watch for

• Two bugs found on the way, both fixed here, both live. A Polish folder
  whose aliases held kawa, kawy, kawie, kawę was refused at startup — the
  overlay wanted every translated list to match English's length, which
  contradicts ADR-0025's "ten forms where English needs three". And
  card_kind_help was the one bank whose entries had no id, so no folder
  could ever land on it, which made ADR-0026's complete unreachable. Its
  id is additive, so no schema bump.
• A register decides which strings are words and which are keys, and
  anything it does not name is prose. That default is the safe one for words
  and the wrong one for a key, so translatable_fields_test.dart fails when a
  string that reads like a key is heading for translation — checked by
  deleting an entry and watching it go red, not by assuming.
• An answer held as one of its options' words follows that option.
  fill[].a and bagpick's answer are values, not indices, so translating the
  options alone would leave every one of those cards unanswerable, silently.
  choices[] needs nothing: it carries correct on the choice itself.
• Two judgement calls a reviewer could take differently. latin is left
  English on the grounds that a Latin name is the same in every language, and
  moduleLabel is translated as prose even though the extractor derives it
  from a module's own words — so a language that translates one and not the
  other could disagree with itself. Neither is ruled anywhere.
• The tool counts 4,654 pieces of prose for a language; the map's own
  estimate is ~4,700, which is the closest thing to an independent check that
  the register draws the line in the right place.

Not here

No language folder ships: this is the tool, not Polish. assets/content/l10n/
stays empty, and the completeness test passes over whatever is there.

Checks

A full round trip against the real banks, outside the suite: 4,654 queued,
filled with marker text, applied, check green — then every written bank
overlaid onto its master through the app's own overlayTranslations without a
refusal. The synthetic folder was deleted afterwards.