Eleven melodic instruments silently refuse presses under the parity sequence at the 14-voice cap

[bdbarnett]

Found 2026-09-03 by the post-tag replay agent (board session), as a side measurement, and reported rather than explained away. Pre-existing: identical counts on the tree before and after the af837de replay, and two of the eleven are outside the 40 instruments that replay touched.

Method. The CPython twin refuses a press silently when no channel is free (if victim is None: continue, no counter; the C side is src/synthio/__init__.c:425). The agent wrapped press with that exact predicate and re-ran the instrument parity sequence (tests/parity/instrument_sequences.py) over all 93 modules through pydevices-audioif at ec3c67f (v0.2.0, cap 14).

Result. All ten drum kits: 0 refused (cr78 and tr808 at exactly 14 permanent Notes included). Eleven melodic modules refuse:

solina 22   jp8000 14   virus 14   b3 10   vp330 10   farfisa 6
k2600 2     polysix 2   ppg_wave 2   prophet_vs 2   vox_continental 2
                                            (86 of 8672 presses)

What it means is the accuracy session's call, not the board's: either the parity sequence is deliberately denser than any real playing (in which case the sequences are exercising refusal, and the gate's goldens include refused notes as silence), or these instruments allocate more voices per note than the cap affords for ordinary chords (solina and vox_continental are ensemble/organ voices, which is suggestive). Either way the refusal is silent, which is the failure nobody hears.

Related: audioif#14 (the ceiling: 14 is what the library was built to), audioif#27 (no gate can see a cap change), the replay report on audiocomponents#2.

[bdbarnett]

Accuracy session's first read, recorded here so the issue carries it: all ten drums refusing nothing while eleven melodic modules refuse points at per-note allocation, not at the sequence — drums are one Note per voice on fixed circuits, the melodic instruments press 2–3 Notes per key (39 of 43 do), and solina at 22 / jp8000 and virus at 14 are exactly where a string ensemble and a supersaw stack most per key. If so, the 14 cap divides by notes-per-key and those instruments run out of keys before the pattern runs out — audioif#14's "the ceiling divides" shown in real material. Then it is neither a bug in the modules nor an unfair sequence but a polyphony product decision. farfisa and vox_continental sitting outside the 40 already rules out af837de. Owner: accuracy session, not tonight.

[bdbarnett]

Audit 2026-09-04: partially addressed — the condition this issue names is unchanged here.

Done upstream: audioif moved the ceiling to 64 at all five sites (8f8b10d), and tests/test_voice_ceiling_consistency.py now asserts they agree.

Remaining: none of that reaches this repository yet. AUDIOIF_PIN is v0.2.0 ec3c67f, 8f8b10d is after that tag, and this repo's environment still reports max_polyphony = 14. So the eleven instruments still refuse presses under the parity sequence exactly as filed.

This closes when the pin moves, not before — and audioif#31 tracks what moving it costs.

[bdbarnett]

Context from the pin move, since this issue's premise has moved: the voice ceiling is no longer 14. It was raised to 64 in audioif (dda8a77 + 8f8b10d), Brad listened to the A/B and blessed it — "an improvement" — and effects Phase 1 moved AUDIOIF_PIN onto the node commit that carries it. Both boards now run firmware built from that line.

So the eleven instruments this issue reports as silently refusing presses under the parity sequence should no longer refuse at all: the measured cause was refused voices, not the limiter knee. On the gate's C3/C4/E4/G4 chord, farfisa admitted [5,5,4,0] channels at 14 — the G silent — and [5,5,5,5] at 64.

Not closing it, because nobody has re-run this specific check at the new ceiling and I would rather it stay open than be closed on inference. What it needs is one run of the parity sequence against the current pin; if the eleven now sound, close it citing that run.

Related: #24's MicroPython leg was red by design until the pin moved, which it now has.