[korexil]

Prompted by a reply on the Hugging Face forum.

• First audio in the four-stage runs is the question event to the server sending its first audio frame (the frame event after the Pacer), not the first sound at the speaker; the board-side delay is not measured. README (all three languages) said end of question to first sound.
• That first frame waits for more than tail_hold_ms (1.2 s) of audio, or for the first segment to end, because each segment's tail is held back to trim trailing silence. The config comment said tail_hold_ms does not affect first audio; it does.
• Gaps are estimates (server-side Pacer with a 0.3 s prebuffer, or the simulator's 0.5 s player); the README footnote now says the 6/6 vs 0/6 result comes from the simulated player.