[bdbarnett]

Fixes #310. A biquad on a low shelf or a low-frequency filter no longer settles at a DC offset in silence.

synthio.Biquad, and so audiofilters.Filter, Note.filter, Echo.filter, freeverb's filter and the graph's LPG node, now run CircuitPython main's arithmetic at ae47a3f15d (adafruit/circuitpython#11679, which also carried #11678):

• 24546b7d1d: fast_sincos reflects the angle into the first quadrant first.
• 0a7df85004: each filter's coefficients get a shift of up to 30 bits, and the state keeps 12 fractional bits in a 64-bit sum that rounds to nearest.

Both are taken in upstream's words. The Python filter chain behind Echo.filter and freeverb on CPython also had PEAKING_EQ's b2 sign the wrong way round, and that's fixed too.

What it does

#310's repro is a 200 Hz LOW_SHELF at A=1.0, quiet noise then silence, through audiofilters.Filter at 36 kHz. Here is the offset over the last 2000 silent samples:

build	before	after
unix MicroPython	+371	0
CPython	+371	0

Upstream's four biquad tests (synthio_biquad_dc_offset, _low_frequency, _high_frequency, _peaking) now print upstream's expected output on both builds. They failed before.

A harness built from upstream's own Biquad.c at ae47a3f15d agrees with this port on 20,000 random filters across all seven modes and rates from 8 to 48 kHz. That is 40,960,000 samples with no difference. The previous kernel fails the harness.

Gates

• verify_dsp with MicroPython and CPython: 62 comparisons, 0 failures.
• verify_graph with MicroPython and CPython: all four probes agree to 0 counts. A planted lpg fault fails it.
• verify_effects, verify_streaming and verify_mixdown_knee pass unchanged.
• The unit tests that reach biquads, filters, echo, freeverb and the graph: 343 tests, OK.

What moved, and why

No CircuitPython release has #11679 yet, so the 11.0.0-alpha.2 oracle now disagrees wherever a synthio.Biquad is heard.

• verify_dsp oracle skips. echo_filter_probe and freeverb_filter_probe matched the oracle byte for byte before this change, and they now skip the CircuitPython comparison. The note on biquad_component_probe is updated to match. MicroPython and CPython are still compared.

• biquad_component golden. It was re-captured because the filter output changed: d9950b16… became ba714db8….

• synthtools_acceptance golden. Its stdout was re-captured. lead, bend and bass run a resonant low-pass and moved, and the other five lines are unchanged. circuitpython_stdout still holds alpha.2's answer, and its note now says why the two differ.

• test_cpython_audiograph_nodes Overdrive comparison. The reference now gets tone as float32, the way the node holds it. With 30-bit coefficients, a tone that isn't float32 moves the output by one count.

• test_cpython_audiobiquad fixed-point test. It now pins the new numbers. A DC burst into a low-pass settles at exactly 0 at 400 Hz and 1000 Hz. At 100 Hz and 40 Hz it settles 1 and 4 counts away, where Q15 used to hold 2631 and 16143.

• Float-precision CI bounds. The single-precision MicroPython cell already allowed biquad_component_probe and echo_filter_probe to differ from CPython on 4 lines each, because synthio.Biquad derives W0 in mp_float_t (synthio.Biquad derives W0 in mp_float_t while the CPython extension uses the shared double helper, so a single-precision target runs a different filter #101). The 30-bit coefficients now carry more of that float rounding than Q15 did, and CI measured 35 and 6 lines. The bounds are set to those numbers, so the gate still fails if either one grows.

When the oracle pin moves to a release that has #11679, drop the three skips and re-capture both goldens from the oracle. upstream-sync.md has the details.

Held for the next resync

These upstream changes landed between alpha.2 and ae47a3f15d and are not taken here:

• the synthio.MidiTrack commits 679e9700ab, ff05e5b620, aad79fe5eb, 6dfa7e5f20 and a54b951de6
• the new audiotap module (e9b2fd49f5, f889146e81)

Downstream digests that run a synthio.Biquad, or the graph's OVERDRIVE with a knob that isn't float32, will move when they pick this up.