[bdbarnett]

Four node kinds for audiograph, so the rest of the polysynths can run as one C graph: Poly6's phase mode, glide and portamento, the LFO shapes their LFO modules use, and sample-and-hold. They are on all three bindings, and no node file outside audiograph changed.

node	what it is	held to
PHASER	audiobiquad.AllPass on one channel, through the same kernel: stages set at build (arg), with frequency, feedback and mix	AllPass, 0 counts, fed the same settings each 256-frame block
LAG	a glide on a pitch CV, the bend on top: a constant-time line (off, on, legato only) or a one-pole	vrack's dsp/glide.py and dsp/utility.py GlideCore, 0 counts and 1 count
LFO_SHAPE	a sine, a falling saw, a square and a random level beside the triangle, all on one phase	vrack's LFOCore's shapes on that phase: sine 1 count, saw and square 0; the triangle is byte-identical to main
SAMPLE_HOLD	latches at every rising clock edge, with an optional lag; with no clock, refreshes at a rate through audioshaper's kernel	SampleAndHoldCore at 0 counts, 1 with a lag; audioshaper.SampleHold at 0

Across interpreters, tests/parity/graph_nodes_probe.py renders all four and verify_graph.py now compares it beside the runner's probe. CPython, MicroPython unix and CircuitPython unix agree on all 307,200 samples, and the runner's probe still agrees on its 41,221. Each of the four new planted faults (a phaser frequency a block late, a legato flag flipped, a shape switch a block late, a clock edge a sample late) makes the comparison fail, and CI runs them first.

Faults planted in the C, one per node, each failed tests/test_cpython_audiograph_nodes.py and were then reverted. They were: the phaser's mix never configured, the glide's ramp one step short, the saw rising, and the hold latching on falling edges.

Also worth knowing:

• The LFO's phase adds a float step each sample, as the triangle always has, and runs up to 0.06% off its rate. That is 0.0065 of a cycle over ten seconds at 1 Hz, where LFOCore stays within 5e-5. Fixing it would move the triangle, so it stays, and docs/audiograph.md says so.
• audiodsp_graph.c now includes audiodsp_fp_contract.h, as every other float file in src/shared does. On x86-64 nothing moves (both probes are byte-identical to main). On a target that fuses multiply-adds, the graph now renders what the desktop does.