[bdbarnett]

CircuitPython's audio outputs play samples, and a sample is filled from C, so there was no way to play audio that Python code or a graph of nodes renders as it goes. This adds audiopump.Ring to the CircuitPython build: a sample your code writes PCM into, which audiobusio.I2SOut, audioio.AudioOut, audiopwmio.PWMAudioOut and audiomixer pull in C, with no Python in the pull path.

from audiopump import Ring
ring = Ring(sample_rate=22050, channel_count=2, frames=256, capacity=8)
ring.write(first_blocks)
i2s.play(ring)
while True:
    if ring.space() >= len(block):
        ring.write(render(block))

It has MicroPython's audiopump.Ring surface (write, space, level, stats, clear, deinit, the audiosample fields; from_usb stays MicroPython's), so one producer runs on both. An underrun plays a whole block of silence, keeps the partial block and counts it in stats()[3]; a write that doesn't fit is short, in whole frames, and counts an overrun.

What differs from the MicroPython twin, and why:

• There is no pump lock on CircuitPython, and a few ports (stm's PWMAudioOut) pull from an interrupt. So clear() doesn't write the consumer's indices: it posts one word saying where to drop to, and the next pull carries it out. space() and level() count the dropped room at once.
• An output that plays the two channels from separate pulls (two DACs on one sample) gets one block per pair of pulls, the right channel offset by a sample, the way audiomixer does it.

It is a new module under src/circuitpython_spike/ and one word in circuitpython.mk. No existing node changes, and nothing in CircuitPython is patched.

Tests. tests/circuitpython/ring_through_mixer.py checks the ring alone bit-exact (with the digests), underrun, overrun, clear(), a wrap many times round a ring whose capacity is not a power of two, and then pulls it through audiomixer in five shapes (mixer buffers that split and span ring blocks, mono, mono into stereo) with a starved stretch, checking every frame's left and right value and that the silence the mixer handed out is the silence the ring counted. 1063 checks pass on a CircuitPython 11.0.0-alpha.1 unix build. Three faults planted in the CircuitPython Ring each fail it: channels swapped, every seventh block dropped, an underrun not counted. The same file passes on MicroPython's unix build, and CI now runs it there (clean-build.yml). tests/test_binding_parity.py holds the two Rings to the same names, and fails when one is removed from the CircuitPython binding.