[repeat98]

• What was wrong or hard: Machines that render on the ColdFire had no admission. The frame guard only drops renders once the frame interrupt is over its share, so a musician with too many machines heard dropouts (or met a DSP stall) and never a refusal, unlike DSP effects, which are refused with a box.

• What this changes: The base bills machines the way it bills DSP effects, with declared prices, not live measurement.

  • A machine's price is reserved[0] of struct modwerk_machine, in percent of a frame (no ABI change); a machine that names none costs 12, a DSP-voice machine 0. The allowance for added machines is 45 % of a frame.
  • A pick over the allowance is refused with a CPU FULL box. A loaded Part that asks for too many keeps the later tracks (in track order) on the silent source and says so.
  • npm run device -- cpu reads the bill and the longest render per track on a dev twin; cpu:PERCENT sets the allowance for a test. A refusal reaches diagnostics as incident cpu, and diagnostics:pull prints it.
  • New beta base 5ef4d447; the five base-bound packages are relinked to it, their code unchanged.

• How I checked it: npm run pr:ready says Ready on main 28a2fc312 (host tests cover the bill, the refusal, the box and the diagnostics record). On the owner's MKII, RAM-booted dev twins of this branch (e609a796; dfbf58d6 with a temporary trace that is not in the branch), OS 1.40C:

  • The owner tried the CPU FULL refusal on the unit with FM Synth on several tracks; the trial counted 6 refusals.
  • Stock load of one busy project: 23 % mean of a frame fresh after boot, 45-57 % at rest after the first PLAY, 73-77 % playing with three modules installed; one stock FLEX render is 4.8-9.5 % of a frame per call.
  • Not done: the 45 / 12 constants are estimates and were not calibrated; POLY8 and VECTOR were not measured; the release base 5ef4d447 itself was not run on a unit, only its dev twins. One run with FM Synth, POLY8 and VECTOR installed and an 8-track pattern ended in a DSP stall ("frames stopped", frame max 163 %, the guard skipped nothing): the budget does not prevent that case yet, see the next point.
  • Known limit: FM Synth 0.1.2's package patches the STATIC renderer entry itself, so with it installed every track renders through its own wrapper, outside the base's dispatch. A refused pick works, but the frame guard cannot shed an FM Synth voice and a refused track is not silenced. The fix belongs in FM Synth (FM Synth 0.1.3 (draft): cheaper on tracks that don't use it #1473) and comes next.

• What the guides lacked or got wrong, and where I fixed it: they did not cover a CPU price or how to read it on a unit; fixed in two guides, three traps are noted but not fixed here.

  • docs/MODULE_PERFORMANCE.md: the CPU price and how to read it with cpu.
  • docs/OCTATRACK_DRIVING.md: the cpu step.
  • Not fixed here: modwerk_events.h is generated from events.json, which nothing says at the header; octatrack:module-packages --base-only skips Sidechain's lagging row, which then needs the new base identity by hand; check_mcf54455 reads a switch's jump table as instructions (worked around in status.c with no-jump-tables on one function; the check wants its own Fix:).

🤖 Generated with Claude Code