`acoustickit` cannot be scheduled

[bdbarnett]

A strike there is a C retune with no frame on it, so schedulable is False and
the kit plays live only. Closing it needs a frame-stamped entry point in the
C node.

From #91.

[bdbarnett]

The audiodsp half is up: PyDevices/audiodsp#138 adds audiopump.STRIKE and audiopump.CHOKE, both targeting audiomodal.Bank.

STRIKE takes the mode table as an array('f') of frequency, decay and gain per mode, validated and flattened when the event is scheduled so the apply half is a loop of stores — no allocation, no raise, no type lookup on the pump thread. CHOKE is the frame-stamped half of clear(), for the hat.

The call this side needs, replacing load()'s direct set_mode calls inside strike():

# on the interpreter thread, when the strike is scheduled
table = array('f')
for index in slots:                      # the slots load() writes today
    frequency, decay = flat[index]
    table.extend((frequency, decay, gain_for(index, velocity)))

queue.at(frame, audiopump.CHOKE, hat_bank)         # hat only, and FIRST --
queue.at(frame, audiopump.STRIKE, hat_bank, table) # same frame applies in
                                                   # schedule order

The layer notes and the stick already ride PRESS/RELEASE, so with these two the whole strike carries a frame and schedulable=True stops being aspirational.

One decision left on this side, and it is a real one. load() writes only the slots a voice uses and leaves the rest alone; STRIKE silences every row the table does not reach, because that is set_modes()'s existing rule — a shorter table is a smaller drum, not a chord of two. Where voices share slots that is the same behaviour. Where a voice is meant to ring through another voice's strike, it needs a bank of its own, which is what the hat already has and why it has one. Worth checking the main bank's voices against that before flipping the flag: if any of them overlap in time by design, the fix is another bank rather than another op.

Not changed here — this issue's audiocomponents half is yours, and the tree is another lane's this sweep.

[bdbarnett]

Proven on a unix MicroPython build, 2026-09-22 — the audiodsp half works. PR PyDevices/audiodsp#138.

strike       ok   two runs of one schedule: 916e56f32080cda1 / 916e56f32080cda1
strike late  ok   peak over the 4 block(s) before its frame=0, over the 4 after=22504
strike choke ok   ringing peak=21598, after a scheduled CHOKE=0

A strike held to a frame is silent until that frame and rings on it; a scheduled CHOKE empties a ringing bank. --fault early moves the schedule to frame 0 and the run exits non-zero.

One thing the build taught me that matters for this side. My first probe fed the bank silence and set gains — and measured nothing, because a modal bank rings what is fed into it. The strike is the gains and the excitation arriving together. So when strike() here becomes schedulable, the excitation has to be scheduled with it (or be continuous); scheduling the STRIKE alone against a source that has already gone past will retune a bank that has nothing to ring.

The rest of the call is unchanged from my earlier comment, and the open decision on this side still stands: STRIKE silences every row the table does not reach, so a voice meant to ring through another's strike needs its own bank, as the hat has.

[bdbarnett]

Arthur: the engine half is in — audiodsp#138 (merged) gives audiomodal.Bank frame-stamped STRIKE/CHOKE ops on the pump's event queue (silent before the frame, peak 22504 after, a scheduled CHOKE taking 21598 to 0 on the unix build). What is left is the instrument side: acoustickit becomes schedulable by emitting those ops. That is a small audiocomponents change for the next sitting.

[bdbarnett]

PR: #104 — the kit is schedulable and it emits both ops.

load() fills a mode table for the whole bank and hands it to Keys.strike_bank; the hat's clear() becomes a CHOKE at the same frame, scheduled first. The live path is byte-identical over a sixteen-hit bar covering every family and the hat choke, proven sensitive against a planted 0.01 % decay change.

Your open decision is answered by sending the whole bank every time. STRIKE silences every row the table does not reach — writing frequency 0, which takes the recursion with it — so a short table would be exactly the defect _silence() was written against. load() already swept the whole bank to mute the other voices while keeping their poles; it writes that same sweep into an array now instead of calling set_mode 70-odd times. So no voice is cut by another's strike, and no second bank is needed.

Your warning about excitation was the thing that made the gate honest. The three new cases in scheduling_seam_live.py all strike a tom, because only kick, snare and sidestick carry layer notes played into the mixer, and the burst is not in the mixer at all — so a tom sounds if and only if a STRIKE reached the bank, with the stick arriving at the same frame.

ok   kit        loudest sample before block 20: 0 (want 0), after: 6981
ok   kitlevels  tom at 127 peaks 6981, tom at 32 peaks 2072
ok   kitchoke   the open hat peaks 5055, the tail 8 blocks after the closing hat peaks 0

kitlevels is the one that could not pass before. With the retune left on the interpreter thread and everything else scheduled — the old behaviour exactly — the loud hit came out at 666 and the soft one at 808: the bar inside out, both ten times down.

All seven cases green; all seven faults fail their own case with a clean control.

Not merged, and one thing is owed: AUDIODSP_PIN is 4e58091, which predates audiodsp#138, so CI's engine has no STRIKE. The same shape as the cebb7ca move for audiomodal.Bank. Moving the pin is its own change. Until then _bank_ops() refuses with a sentence and the kit plays live exactly as before.