Detect click/pop discontinuities and correlate them with playback events

[engkimo]

Problem

A caller can hear a click or pop when playback is interrupted, a provider chunk boundary is mishandled, or adjacent PCM frames have a large discontinuity. VoxBench records queue clearing, provider chunks, and stage PCM quality, but it does not yet emit a bounded click/pop observation.

Goal

Add a deterministic, explainable discontinuity metric and correlate suspicious points with existing playback, barge-in, and serializer evidence.

Design requirements

• Define the signal contract first: sample format, window, threshold units, minimum signal level, and treatment of silence.
• Prefer a simple explainable time-domain detector before a learned classifier.
• Report bounded metadata only; do not persist additional raw PCM.
• Keep stage media time separate from wall-clock event time.
• Avoid duplicate incidents across adjacent windows.
• State confidence and preserve remote_playout_observed: false unless another adapter proves playout.

Acceptance criteria

• Synthetic fixtures cover a clean boundary, a zero-crossing boundary, a strong non-zero discontinuity, silence, and clipped PCM.
• The observation includes stage, media-time location, magnitude, and detector contract/version.
• Existing barge-in or queue-clear evidence can be referenced when temporally correlated.
• The default test suite is deterministic and requires no audio device or provider key.
• Documentation explains what the detector can and cannot prove.

Likely areas

• src/voxbench/observability/observer.py
• src/voxbench/telephony/audiosocket.py
• src/voxbench/control_plane/run_api.py
• tests/test_observability.py
• tests/test_audiosocket.py
• Web incident rendering if a new incident is added

Out of scope

• Claiming subjective audibility from a single threshold.
• Retaining caller recordings by default.
• A provider-specific implementation.

Audio/DSP and realtime review are especially welcome on this issue.