[FEATURE] Desktop dictation: revise the live transcript while speaking instead of correcting only on Enter

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Dictation in the desktop app streams text into the composer while I speak, but the streamed text is never revised during the recording. Misheard words stay on screen, uncorrected, until I stop dictating or press Enter. Only then does a final pass fix them.

Other real-time speech-to-text engines work differently. As more audio arrives, they keep rewriting the unstable tail of the transcript, so a misheard word gets fixed within a second or two, while you're still talking.

Right now, the live text gives the wrong signal:

1. While I speak, I see a mistake and have to decide on the spot whether to stop and fix it or keep going and hope the final pass catches it.
2. After I press Enter, I have to read the whole message again to see what the final pass changed and whether it fixed the mistakes I noticed.

So one dictated message takes two passes: speaking, then proofreading. The live preview is supposed to let me skip the second pass, but here it causes it.

Proposed Solution

Revise the transcript live, the way streaming STT does:

• Treat recent words as provisional and keep re-decoding them as more context comes in, replacing them in place.
• Freeze text once it's stable (for example, after a pause or a sentence boundary), so settled text doesn't keep moving.
• Optionally, style provisional text differently (dimmed or underlined) until it's final, so it's clear which words may still change.
• Whatever is on screen when I stop should be what gets sent. The final pass shouldn't make surprise changes at send time. If it does change something, briefly highlight the change.

Alternative Solutions

• Show no live text at all, only the final transcript. This is worse for feedback, but at least it's honest.
• Run the correction pass when I stop dictating instead of on Enter, and highlight the words it changed, so I can check them before sending.

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

1. Click the mic in the desktop composer and dictate a long prompt with technical terms.
2. A term is misheard early on, and the wrong word stays in the composer for the rest of the recording.
3. I either break my train of thought to fix it, or finish talking and press Enter.
4. On Enter, the text is corrected, and I have to re-read the whole message to check what changed before trusting it.

With live revision, the word would be fixed a moment after it appeared, and what I see while talking could be trusted.

Additional Context

• This is mainly about the desktop app's dictation. The CLI voice mode likely has the same streaming/finalization split.
• Related, but not the same: #95510, where the finalization pass rewrites a correct live transcript into a different script. That issue shows the same design, where the streamed text and the committed text come from separate stages. This request asks for the streamed text to be revised continuously so the two stages agree.
• #39239 (closed) asked for context-aware transcription accuracy in general. This request is about when corrections appear, not how accurate they are.