TT-7596 TT-7609 fix: tag PBT takes with their language and load an existing take mid-navigation

[nabalone]

Problem

Takes uploaded untagged (TT-7596). The media upload payload sent the language attribute as 'language-bcp47'. The API expects the single lowercase token languagebcp47 — unlike the other multi-word attributes it is not dasherized — and silently drops the unknown key. Every Phrase BT take was therefore stored with no language.

Anything that scopes rows by step language then failed to match them:

• the PBT Transcribe step reported "Some Phrase Back Translation segments are missing recordings" (TT-7596);
• navigating back to an already-recorded segment showed an empty recorder, because recordingRow — and so MediaRecord's mediaId — never resolved.

Both cleared once the record was re-read from the server, which is why leaving the step and returning "fixed" it.

Load dropped mid-navigation. On top of that, MediaRecord loaded an existing take only on a mediaId change, and skipped it outright while a load or save was in flight (if (!loading) handleLoadAudio()), with nothing to retry it. Changing clause during that window lost the load until the step was remounted. An in-flight load was also never abandoned, so a late-resolving fetch could hand the previous clause's audio to the clause just navigated to.

Changes

• store/upload: send languagebcp47; the type carries a comment about the non-dasherized key so it does not regress.
• MediaRecord: defer the load instead of dropping it — a new effect keyed on [loading, mediaId] runs the skipped load once the recorder is idle, guarded by the id a load was started for so it runs at most once per take.
• MediaRecord: handleLoadAudio captures the id it was started for and bails at each await if mediaId has moved on, so a stale fetch cannot paint the wrong take.
• MediaRecord: a just-uploaded id is marked as already present, so saving no longer blanks the take and re-fetches it.
• The reset() stays tied to a real mediaId change (not to loading), so a take whose upload failed is not wiped out from under the TT-7583 retry.

Test plan

• Manually tested in the app by the author — recording successive Phrase BT segments and returning to an earlier one now shows that segment's take.
• New MediaRecord.load.test.tsx. Verified failing-before / passing-after against unfixed source: pre-fix the mid-load case never loads the second take (loadCalls stays at the abandoned url) and hands the stale blob to the waveform; post-fix both assertions hold.
• npx jest src/components/MediaRecord → 12/12 pass, including the pre-existing suite (the TT-7583 "keeps save available so the same take can be retried" test guards the reset behavior above).
• tsc --noEmit clean; Prettier clean on the touched files. actions.tsx has a pre-existing formatting deviation that also fails on develop, left alone to keep the diff to one line.

Note: the cypress-tests check failure is unrelated to this diff — the Cypress binary is missing from the runner cache, so no spec runs. It is failing the same way on other current PRs and on recent develop pushes.

🤖 Generated with Claude Code