[hsliuustc0106]

What

docs/speech.md only — no code, no behavior change. Two updates from the 2026-10-08 quiet-machine measurement session:

1. Performance & verification gains the five-checkpoint table (first-audio p50 / sustained RTF p50 / peak, one session, one machine class) plus the coexistence line (speech RTF 0.47 with a 0.5B chat resident) and the 16 GB → 0.6B guidance.
2. Voice cloning section: the stale "~0.35–0.7 s TTFA — a known optimization target" claim is replaced by the measured 0.28 s (1.7B) / 0.15 s (0.6B) and the tts: compiled bucketed ICL prefill — clone TTFA is 68% launch-bound eager prefill #112 floor verdict — the ICL prefill is compute-bound at the 4-bit GEMM floor (compiled prefill benched identical to eager), so there is no kernel-level headroom on the 1.7B; the 0.6B pair is the clone pick for small machines (peak 3.7 vs 5.9 GiB).

The published v0.1.0 release notes' matching known-limit line was corrected separately (user-approved, via gh release edit — not part of this diff).

Provenance

• All numbers measured 2026-10-08, quiet M4 base · 16 GB, direct serving path (no HTTP), texts/seeds/interval identical to the bench scripts; full logs and methodology in #2 (table + 0.6B A/B), #39 (coexistence re-verify), #112 (clone floor negative result).
• Cross-check: solo RTF 0.403 on the bench_stream_rtf.py methodology, inside the Reproduce the quiet-machine TTS streaming benchmarks on your Apple Silicon (RTF record: 0.37–0.42) #84 cold-morning envelope (0.29–0.45); the historical 0.73–0.90 spreads were parallel-session GPU-contention artifacts.
• Docs-only per the house rules — no A/B required; no tests touched.