TT-7557 fix: scope Transcribe tasks to the step's language

[nabalone]

Problem

TT-7557 — a team can configure a Phrase Back Translation record + transcribe step pair per language (English, Arabic, French). Every PBT recording showed up in every PBT Transcribe step, because TranscriberProvider built its task list from plan media filtered only by artifact type.

Change

The Transcribe step's configured language (settings.language) now scopes the task list. When it is set (and not und), only media tagged with that bcp47 become tasks.

• matchesGuidedOutputRow.ts — extracted the language check into mediaMatchesStepLanguage(mediafile, stepLanguageBcp47); matchesGuidedOutputRow delegates to it, so row-level and media-level filtering can't drift.
• TranscriberContext.tsx — TranscriberProvider takes an optional stepLanguageBcp47 prop and applies the filter to getMediaInPlans(...) before planMedia / planMediaRef are set, so everything downstream (addTasks → rowData, expanded groups, auto-selection, allDone, the Transcriber pane) only sees same-language media. The prop is in the effect deps, so switching steps re-derives the list.
• PassageDetailTranscribe.tsx — passes the stepLanguageBcp47 it already computes from step settings (the same value driving phraseRegions and hasBtRecordings).

Filtering at the source array rather than at render matters: rowData also drives auto-selection and the Transcriber's own row lookup, so filtering in TaskList would hide rows while still letting the pane load and save them.

Vernacular steps are exempt (second commit)

A vernacular Transcribe step has no step language — it transcribes in the org vernacular, and its media carry no languagebcp47. If such a step's settings still held a leftover language value (e.g. it was once a PBT step), the filter would scope its task list to nothing. stepLanguageBcp47 now resolves to undefined without an artifactTypeId; every other consumer of it already bailed in that case, so that memo is the single place to decide it. Raised by the Devin review on this PR.

Notes

For non-vernacular artifact types the match is strict, so a recording with no languagebcp47 is dropped when the step has a language. That's the rule hasBtRecordings already applies to gate the whole pane, and the record step's planLegacyPhraseBtClaim stamps untagged legacy outputs with the step language.

Whole Back Translation steps also offer a language picker, and WBT recordings don't appear to be stamped with a language. That combination is already hidden by the pre-existing hasBtRecordings gate before the task list renders, so this PR doesn't change the visible outcome — but whether WBT should be language-scoped at all is worth a separate look.

Test plan

• npx jest src/components/PassageDetail — 34 suites, 217 tests pass, including new matchesGuidedOutputRow.test.ts coverage for mediaMatchesStepLanguage (unset/und step language, matching tag, other language, untagged media).
• npm run typecheck clean.
• Manual: on a team with per-language PBT step pairs, record PBT audio in two languages for the same passage and confirm each PBT Transcribe step lists only its own recordings, and that the initially selected task belongs to that step. Also confirm a plain vernacular Transcribe step still lists its audio.

🤖 Generated with Claude Code