audiofilters.Phaser holds −4 LSB of DC forever — its own int16 all-pass state, not the fixed-point biquad of #23

[bdbarnett]

Split out of #23, which its title currently carries. They are not the same defect and #23's fix will not touch this one.

What is measured

audiofilters.Phaser at its defaults (6 all-pass stages), 48 kHz stereo, 0.09 s of 1 kHz at −6 dBFS then digital silence, pulled to 600 blocks: the output settles after 20 blocks to a constant −4 LSB and stays there. Re-run to 3000 blocks (32 s of audio) with the settle window excluded, the set of distinct sample values over blocks 60..3000 is exactly [-4]; block 100 and block 2999 are identical. This is a held DC constant, not a decaying tail.

Rate and channel count matter and must be stated with any re-measurement: the same probe reads exact zero at 22.05 kHz mono. A run at another rate will look like the defect is gone.

Full table, method and harness note: the effects-program Phase 0 comment on #23 (#23 (comment)).

Why it is a different defect

audioif_phaser.c never calls audioif_biquad. grep -rln audioif_biquad src/ matches only:

src/synthio/Biquad.c
src/synthio/Biquad.h
src/shared/audioif_biquad.c
src/shared/audioif_biquad.h
src/shared/audioif_trig.h
src/cpython/_audioif.c

#23's cause is the AUDIOIF_BIQUAD_STATE_SHIFT 12 recursion in audioif_biquad.c:159-182 — state carrying 12 bits below the sample LSB, round-to-nearest at every step, no dither and no leak term. The Phaser's hold has a different origin: its own int16_t allpass_words[] / feedback_words[] (audioif_phaser.c:28-38), which are in plain sample units with no fractional bits at all. Fixing the biquad's state format will not move the Phaser's −4 LSB by one bit.

What is not proposed

audioif_phaser.c is a CircuitPython-ported kernel. It is never modified — the standing rule, and Brad's hold of 2026-09-03 on #23. This issue exists to record the defect against the right kernel, so that:

• Three shipped filter configurations never return to silence — Phaser defaults hold −4 LSB of DC forever #23 stays about the fixed-point biquad and its title stops implying the Phaser is downstream of it;
• the effects program's Gate 0 answer can cite the right cause. Its node candidate 9 (a DC-clean float biquad and a first-order all-pass in an audioif-own module) carries the all-pass kernel precisely because this row cannot be fixed by fixing Three shipped filter configurations never return to silence — Phaser defaults hold −4 LSB of DC forever #23;
• the Phaser class's Tier 1 invariant ("a decaying tail reaches exact zero — no held DC") has a filed cause to point at rather than a rediscovery every time someone measures it.