Add a wake-word acknowledgment chime

[jamditis]

Play a short chime the instant the wake word fires, before capture, to signal the mic is now listening for the command, then return to passive wake-scan after the turn.

Implement in the live-loop wrapper (audio.py), fired right after stream_detect_wake returns and before capture_request, not in core pipeline.py. A 16 kHz mono int16 two-tone rising cue with 10 ms edge fades avoids clicks; a prototype generator already exists.

[jamditis]

Tonight's live run argues for making this two-stage, and for keeping both stages model-free.

Two stages:

1. A short chime the instant the wake word is detected (this issue), before capture.
2. A brief "working" cue once the handoff to the brain succeeds — fire it when bot-spren send returns 0, i.e. the message actually landed in the inbox.

No LLM for either cue. These are state transitions, not language tasks. A model would add cold-start latency, nondeterminism, and cost to a cue that must be instant and identical every time. Generate the tones directly.

Rationale to record: tonight we repeatedly could not tell "it didn't hear me" from "it heard me and is thinking" because the wake fired silently. The wake chime answers the first; the post-handoff cue answers the second.