I2SOut.starved() reads thousands of bytes on a clean P4 playback, and a gc.collect() makes it go down #26

On the ESP32-P4 panel (ESP32_GENERIC_P4 PRE_REV3_C6_WIFI, MicroPython 1.29), I2SOut.starved() doesn't work as an underrun count. Through audiodev's audio_out() I played a looping 1 kHz RawSample at 24 kHz, with no recorder open:

play alone 1 s:                9654
3 s:                           9689
gc.collect() (4 ms):           510
after 3 more gc.collect():     1536
stop, play again, 1 s later:   9658

Starved audio can't un-happen, so the count should never go down, and a clean playback should read 0. An acoustic recording of the same kind of playback (through the panel's microphones, with the change in #25) shows no dropout at all: no 2 ms window falls below half the median envelope in six seconds.

So the number moves, but it isn't measuring silence on the wire. My guess is the wall-clock witness (due = fed0 + elapsed × byte_rate) set against fed0, which is taken at the last armed write. I haven't confirmed that. Until it's fixed, the honest underrun witness on this board is the recording, or i2s_sink_bytes() − i2s_dma_bytes() sampled over time, which stays flat to within one descriptor when nothing starves.