The legacy I2S driver buffers i2s_write() data in a TX DMA ring
(8x64 samples = 64 ms @ 8 kHz) and hands a buffer back to i2s_write()
only after the DMA finished playing it. After the last queued event
the sound task called i2s_zero_dma_buffer(), which memsets the whole
ring instantly — erasing everything not yet played. Any sound <= 64 ms
was written into free ring space without ever blocking, so the zeroing
wiped it before the DMA reached a single sample: completely silent on
hardware. Longer tones force i2s_write() to block on a full ring,
pacing rendering to real-time playback — hence only long beeps were
audible, and the fade-envelope cap never moved the threshold. Every
game beep <= 80 ms (link-active 10, laser 30, sever 40, bleep/ICE 50,
feedback 80) sat in the dead zone. The emulator never showed it: its
i2s_write shim fed SDL directly and had no ring at all.
Fix: replace the instant zeroing with a paced silence purge — once the
queue drains, write one ring + one chunk of zero samples through
i2s_write(), reusing each DMA buffer only after the DMA played it. The
sound plays out fully and the ring then loops silence (still kills the
tail-replay bug the zeroing existed for). Keep the envelope cap at
total/4 so short tones keep a steady-state portion.
Emulator shim gains a TX DMA ring model (write/play cursors + bounded
DAC-emitted log) so the bug class reproduces headlessly: unit/
sound_short_tone_reaches_dac fails on the zero_dma code, passes on the
purge (A/B verified); shim_zero_dma_erases_unplayed pins model fidelity;
sound_long_tone_plays_out_fully guards against truncation. SoundEngine
gains a pumpOne() test seam and begin(spawnTask) for deterministic
driving. Full suite: 26 unit + 7 integration + 4 E2E, 0 failures.