[FANNG1]

Summary

• add local-baseline for 40–60 fixed 60-second audio files (default 50)
• warm up Ray/model/cache outside measured repeats and reuse one manifest across two isolated Lance outputs
• generate JSON/Markdown comparisons with throughput mean, delta, CV, memory summaries, and repeatability SLO
• preserve a baseline summary when a measured repeat fails or times out
• document standard versus small diagnostic runs

Validation

• uv run pytest (117 passed, 1 deselected)
• independent agent review completed; failure-summary findings addressed

Follow-up to #22.