Capture cannot keep up: short-speech queues, previews and final results
bufferOverflow is a failure boundary; partial previews are not complete recording
The short-speech PCM pipeline uses a serial queue and bounded waiting source frames. Exceeding max(ceil(sourceRate / 2), 4096) records bufferOverflow, stops admission and ultimately fails. snapshot does not drain the queue; it captures current output only. Visible previews/levels do not prove complete capture or delivery.
Different from format and route review
Format admission checks shape; route recovery checks system events. Here queue, stop, snapshot and final meanings under pressure matter. This scope is short speech, not long-note limits or hardware performance testing.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Budget | max(ceil(sourceRate / 2), 4096) source frames | Waiting amount/error | Fixed half-second end-to-end SLA |
| Synthetic arithmetic | 48000 Hz → 24000 waiting frames | Frames, not bytes | Exact whole-app RAM cap |
| Overflow | bufferOverflow failure, admission stops | Actual error/incomplete state | Silent drops still count as full success |
| snapshot | Does not drain for preview | Current versus final output | Every queued block included |
| Normal requestStop | Stop admission, process queue/flush tail | Returned final/error | Stop tap proves instant completion |
| Duration | Finite parameter up to 30 seconds | Normal limit stop versus durationExceeded | Every 30-second capture errors or long notes share limit |
| cancel | Clears pending/output and resets | Cancellation versus committed effects | Hidden complete recovery backup |
| Handoff | Final/original guards still apply | Result/target/prior attempt | Tuning or recapture grants resend |
Synthetic pressure result
A preview exists, but final reports bufferOverflow. Do not present the preview as complete. Retain failure/gap evidence without arbitrarily growing queues, bypassing model/format checks or diagnosing hardware capability from one result.
Source budgets are not runtime/noise/accuracy evidence. Processing blocks, output and other allocations are not represented by waiting frames alone. No stress test, recording or pipeline modification ran.
FAQ
Does half-second frame budget mean half-second latency?
No. Waiting admission differs from end-to-end latency/scheduling.
Are 24000 frames 24000 bytes?
No. Frames, channels/layout and total memory differ.
Can visible text justify ignoring overflow?
No. Partial preview does not turn final failure into completeness.
Does snapshot wait for every block?
Not for preview. Distinguish current output from final stop results.
Do all long notes have a 30-second limit?
No. This is short speech; long-note disk recording differs.
Is automatic queue tuning/cloud fallback added?
No. This reviews source failure boundaries, not tuning/fallback/zero-loss guarantees.
Reading and source
Format admission · Route recovery · Recapture gaps · Start · Privacy
