fix(vad): bound live segments to guarantee transcript updates

[safvanatzack]

Problem

The live pipeline sends audio to ASR only when the active VAD segment closes. This leaves live transcript delivery unbounded: continuous speech or system audio can keep one segment open, grow the VAD buffer past its warning threshold, and withhold transcript updates even though audio is arriving.

PR #679 should use a 500 ms live redemption time to reduce the known 400 ms over-fragmentation. That is intentionally separate from this issue: redemption time controls ordinary pause handling; it cannot guarantee a bounded ASR request or a bounded interval between live transcript updates.

Import and retranscription keep their 2000 ms batch policy and are out of scope.

Goal

Give the live path an explicit bounded-segment policy so continuous speech produces incremental transcript updates without reintroducing short-request fragmentation, clipping, duplication, or silent loss.

Proposed policy

• Define a live ASR segment target of 20 seconds and a hard maximum of 25 seconds, subject to replay validation.
• When an active VAD segment reaches the hard maximum, dispatch a safe live ASR segment rather than waiting indefinitely for VAD to close it.
• Preserve speech continuity across forced boundaries: no lost words, duplicated audio, or timestamps outside the recording duration.
• Keep the 500 ms live redemption policy independent from the maximum-segment policy.

Non-goals

• Changing positive/negative VAD thresholds or the 500 ms redemption value from fix(vad): stop fragmenting live speech into sub-4s ASR requests, and filter hallucinations on every engine #679.
• Changing import or retranscription segmentation.
• Changing transcript cleanup, hallucination filtering, or confidence-gate behavior.

Acceptance criteria

1. A continuous-speech or continuous-system-audio recording longer than two minutes receives incremental live transcript updates; the active VAD buffer never remains undispatched beyond the hard maximum.
2. Normal conversational speech retains the reduced short-request fragmentation expected from the 500 ms live redemption policy.
3. A forced boundary preserves complete, ordered transcript text and valid timestamps; stopping mid-utterance still satisfies start <= end <= recording duration.
4. No duplicate transcript rows or silent chunk drops are introduced at forced boundaries.
5. Regression coverage exercises a forced boundary and a final flush; manual replay covers the continuous-system-audio case that exposed the unbounded buffer.

Related work

• fix(vad): stop fragmenting live speech into sub-4s ASR requests, and filter hallucinations on every engine #679 — live VAD redemption and final-segment timestamp correction.
• split shorter vad, make it return transcription as soon as possible #693 — related live VAD work; this issue defines the explicit bounded-delivery policy that remains required.