Per-arm WER budgets from two-run evidence; document streaming variance

Tighten per-arm WER budgets to measured values

Per-arm WER budgets in the golden manifest

Add per-clip WER budgets and specdec acceptance floors to the golden manifest

Add human-reviewed golden samples for Qwen3-ASR functional tests (transcripts + speaker turns; word timings deliberately excluded)

Add two-speaker glitchy-audio clip for ASR/diarization golden review

restructure: existing suite -> dev/ (validation+tuning); new held-out YouTube clips -> test/ (8 sources, ar/it firsts; SOURCES.md provenance)

asr-test-suite: 9-language length-ladder 16k wavs + ladder sources (fetch_test_audios.py provenance)

initial commit