feat(document): add immutable transcript word edits - #540

The auto pass queues every asset in the document. That was harmless while every
asset was footage; imported audio broke it. Measured on this branch, with a
four-minute bed on the music lane: 35 seconds of whisper inference at editor
open, producing 164 segments of transcribed music.

It went unnoticed until now only because `hasLocalSttEngine()` was false on this
machine — the helper was not built, so `stt:transcribe` failed instantly. Staging
the CI-built binary turned an instant failure into real work, and the cost became
visible.

`assetCanCarrySpeech` answers the question from the TIMELINE, because the asset
cannot: `AxcutAsset.kind` only knows `video | audio`, and the voiceover/music
distinction lives on the region. An audio asset qualifies exactly when some region
playing it sits on a voiceover lane.

That rule is stable under a lane change, which matters because both the inspector's
selector and the agent's `setAudio` write `kind`: music -> voiceover queues it (it
is speech now), and voiceover -> music discards nothing, because the transcript
already exists and an asset that has one is skipped. The round trip is lossless
rather than paid for twice.

This is the rule settled in #560; it lands here rather than waiting for that
issue's phase 1, because the freeze hits anyone who drops a music bed and it does
not depend on the transcript work #540 is doing.

Not fixed here, and the bigger half: the pass blocks the UI at all because
`extractMono16kFromVideoUrl` decodes the whole file in the RENDERER before the
helper ever sees it. See the PR discussion.