Early transcription ends a sentence at every piece seam, so a pause mid-sentence becomes a full stop the one-shot result doesn't have

## What happens

When a dictation is long enough to be cut into pieces while the key is held, `PieceJoiner.seamed` (`Sources/UttrflowPipeline/PieceJoiner.swift:50-54`) ends every piece except the last with `endedAtSeam` (lines 57-64), which calls `WordShape.finished(text)` whenever the destination uses terminal stops. The doc comment says it plainly: a seam is "a pause the windowing cut at, which every stopping place reads as a sentence end."

A pause is not always a sentence end. On a real-voice test of early transcription, a sentence with a pause in the middle came back from the live path with a full stop at the seam, splitting it in two, while the one-shot transcription of the same audio kept it as one sentence. An invented example of the shape: "we moved the review. To Thursday because the room was taken" instead of "we moved the review to Thursday because the room was taken".

This behaviour was introduced deliberately by #362 (fixing #183, where seams lost their stops in chat destinations). That fix is right for a seam that really is a sentence end, and wrong for one that is not.

## Why it matters

The live path exists to make the wait shorter, and its output is supposed to match the one-shot result. A wrong full stop in the middle of a sentence is a visible error in the inserted text, it appears only on longer dictations, and it will not show up in any test that cuts at a real sentence boundary.

## How to reproduce or check

- In `Tests/UttrflowPipelineTests/PieceJoinerTests.swift`, `PieceJoiner.seamed(["we moved the review", "to Thursday because the room was taken"], under: .standard(for: ))` gives a first piece ending in a full stop.
- On a real microphone (see #55): dictate one long sentence with a clear pause in the middle, long enough to be cut, and compare with a one-shot transcription of the same recording (retry it from the Dictation page).

Related: #55 (real-microphone verification of early transcription), #230 (whole-dictation policies evaluated per piece).

Start with PieceJoiner.seamed in Sources/UttrflowPipeline/PieceJoiner.swift and run the related tests in Tests/UttrflowPipelineTests/PieceJoinerTests.swift and DictationPipelineSeamTests.swift. Review the existing #362 assertion and agree which signal determines a sentence boundary before changing behavior. Done means mid-sentence seams avoid added stops, genuine sentence ends and the #183 chat cases remain correct, with an evaluation case added if possible.

Written by the indexing model from the issue text.