Add LUFS and true-peak observations with explicit measurement contracts

[engkimo]

Problem

VoxBench currently records RMS, sample peak, full-scale sample percentage, gain, silence, and duration evidence. These are useful for debugging but do not describe integrated loudness, loudness range, or inter-sample true peak.

Goal

Add standards-informed loudness and true-peak observations without turning a single number into an unsupported quality verdict.

Design requirements

• Propose the measurement standard and algorithm before implementation.
• Record sample rate, channel contract, window/gating behavior, units, and implementation version.
• Define how short telephone utterances become insufficient rather than misleading values.
• Keep sample peak and true peak distinct.
• Keep stage-native and comparison sample rates explicit.
• Do not add a mandatory heavyweight dependency without discussing packaging and platform impact.

Acceptance criteria

• Deterministic synthetic fixtures cover tone, silence, clipped signal, short/insufficient audio, and level changes.
• Results are finite, bounded, and use explicit units (LUFS, LU, dBTP).
• Timeline/stage presentation distinguishes measured values from policy thresholds.
• A missing optional implementation reports unavailable or insufficient, not zero.
• Documentation states limitations for 8 kHz telephony audio.
• Python checks and Web build pass.

Likely areas

• src/voxbench/media/
• src/voxbench/observability/observer.py
• src/voxbench/control_plane/run_api.py
• src/voxbench/verification/
• stage detail Web UI and tests

Out of scope

• Declaring one universal “good” loudness threshold for every provider or route.
• Retaining additional raw media.
• Replacing full-reference scorers such as ViSQOL.

Audio measurement expertise and review are welcome.