feat(audio): live push/queue playback stream for on-the-fly synthesis

[InauguralPhysicist]

What

Adds five builtins — audio_stream_open / audio_stream_push / audio_stream_queued / audio_stream_clear / audio_stream_close — a dedicated SDL playback device opened in queue mode (callback-NULL) and fed block-by-block via SDL_QueueAudio. It coexists with the callback mixer (audio_open) and the SDL_mixer music device; the OS mixes all three.

This is the capability the fixed-buffer mixer channels can't provide: audio generated on the fly as notes start and stop, with no pre-rendered clip.

Why (forcing function)

DeslanStudio musical typing (F-DS-17). Sustain-while-held notes need a stream you can feed each tick from the synth's stateful voice pool — the whole-buffer audio_play can't sustain. Rather than hack it in the app, the streaming primitive is forged upstream where every future audio consumer can use it.

Design notes

• push is a pure output sink — float→int16 through the existing audio_convert_samples (same 64 MiB cap / clamp / NaN guards). SDL_QueueAudio copies the block, so the temp buffer is freed immediately. Not trace-recorded.
• queued reads a live, timing-dependent value that steers the caller's refill control flow, so it carries the TRACE_NONDET_TAKE/RECORD tape pair exactly like audio_capture_read: under EIGS_REPLAY the recorded queue depth replays and the refill pump makes the same push decisions, keeping sessions byte-deterministic.
• Registration via the ext_names.h X-macro (zero lint/registrar drift).

Validation

• -Wall -Wextra -Werror clean (gfx + default builds)
• test_audio 53/53 under SDL_AUDIODRIVER=dummy, plus the no-device skip path
• ASan + UBSan clean (detect_leaks=1)
• Full suite 2931/2931; discoverability gate [99b] green (documented in docs/BUILTINS.md)
• 8/8 stable audio runs

🤖 Generated with Claude Code