Verify early transcription on a real microphone, not only in unit tests

## Context

PR #50 transcribes and tidies each pause-cut piece of a dictation while the key is still held. It is covered by unit tests against synthetic audio and an injected clock, but nobody has yet run the new path end to end with a real microphone.

## What to do (after #50 merges)

1. `make app`, open `dist/Uttrflow.app`, hold the shortcut and speak three or four sentences with natural pauses.
2. Confirm the inserted text equals what a single-shot dictation of the same speech produces (use the Dictation page to compare).
3. Try: a long pause mid-sentence, a sentence with a name from your dictionary, releasing the key mid-word.
4. Report the machine, macOS version, engine chosen in Settings, and any divergence.

Anything that differs from the single-shot result is a bug worth its own issue.

Run `make app`, open `dist/Uttrflow.app`, hold the shortcut, and speak several sentences with natural pauses. Use the Dictation page to compare the inserted text with a single-shot dictation, then try the listed pause, dictionary-name, and mid-word cases; done means reporting the machine, macOS version, selected engine, and any divergence.

Written by the indexing model from the issue text.