Add per-row and pooled latency metrics to eval command

[alexkroman]

Instruments the eval command to measure and report transcription latency for each item and compute latency percentiles (p50, p90) across the dataset.

Summary

The evaluation command now tracks wall-clock latency for each transcription request using time.perf_counter() and includes per-row latency in both JSON and human-readable output. Latency metrics are pooled across all rows (including failed ones) to compute p50 and p90 percentiles.

Key Changes

• Latency measurement: Wrapped _transcribe_one() with time.perf_counter() calls to measure request duration; introduced _Timed dataclass to pair transcription outcomes with their latency
• Per-row latency: Added latency field to all result rows (both scored and failed); updated _ItemResult to carry latency alongside words for pooling
• Percentile computation: Implemented _percentile() function using linear interpolation between ranks (numpy's default method) to compute p50 and p90 from latency values
• Pooled metrics: Extended _pooled_metrics() to compute latency_p50 and latency_p90 across all transcribed items
• Output formatting:
  • Added _secs() formatter for latency display (e.g., "1.50s")
  • Updated human-readable table to include LATENCY column when latency data is present
  • Extended summary line to show "latency p50 X.XXs · p90 Y.YYs"
• Test coverage: Added comprehensive tests for latency tracking in both success and failure paths, percentile interpolation edge cases, and output formatting

Implementation Details

• Latency is measured for every transcription attempt, including failed requests, ensuring the distribution reflects actual API behavior
• The _Timed wrapper is immutable (frozen dataclass) to prevent accidental mutation
• Failed rows include latency in their output row dict alongside the error message
• Percentile calculation handles edge cases: single values, odd/even counts, and boundary quantiles (q=0, q=1)
• Test helper _fake_perf_counter() pins the timer for deterministic latency assertions

https://claude.ai/code/session_01SWJQ3VVvR2YLyPDbtrw6tU