[bdbarnett]

audiograph is no longer MicroPython only. The wheel has it (import audiograph on CPython), and the CircuitPython build compiles it, both the unix port and the firmware, through circuitpython.mk's module list. All three bind the same src/shared/audiodsp_graph.c with the same arguments and the same refusals, and no node file changed.

• CPython: a Graph type in _audiodsp.c, re-exported as is by src/cpython/audiograph.py. The constants are read by name from the C enums. A delay's wow table is held through a memoryview, so the array can't be resized under it.
• CircuitPython: src/circuitpython_spike/shared-bindings/audiograph/ and shared-module/audiograph/, laid out like the other modules.

The comparison. tests/parity/verify_graph.py runs graph_parity_probe.py on CPython and on each interpreter you name, then compares every output sample. The patch has two voices with every node kind, groups that park, a fed buffer, moved gains and cutoffs, a delay reset mid-tail, and a mono graph. The tolerance is 0 counts. Locally, CPython, MicroPython unix (double and single precision) and CircuitPython unix agree on all 41,221 samples.

Two planted faults, applied to every interpreter after the first, each make the comparison fail:

• lsb: one fed input sample is one count off, and 615 samples differ by up to 2.
• late: one set() lands a block late, and 3,197 samples differ by up to 183.

CI runs both faults first, then the CPython against MicroPython comparison, beside graph_probe. As with verify_dsp, the three-way that includes CircuitPython is a local run. graph_probe.py passes on all three, with the same digests on each.

Tests. tests/test_cpython_audiograph.py runs the runner's probe on the wheel. It also checks that the module exports every name the board module does, that a borrowed table is held, and that a dropped graph frees its memory; a planted leak fails that last check. test_binding_parity now covers audiograph (22 types), and Graph is exempt from the node deinit rule because nothing plays it.