chore(inference): keep analyze metadata and use relative benchmark audio

scripts/analyze no longer zeroes elapsed time or reuses one request id;
each run gets a UTC-timestamped id. The golden test still passes the
deterministic flag itself. Benchmark manifests now point at
benchmarks/audio/, which is gitignored for .wav files, instead of
/path/to placeholders.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
8e0e45f922flamboh committed on 9/17/2026, 5:13:08 AMparent6cfe5de
6 files changedLine totals unavailable