[repeat98]

• What was wrong or hard: FM Synth 0.1.2 puts its own renderer straight into the tracks' renderer slots (a detour at the frame builder's slot copy 0x4000d514, plus the STATIC table entry). The base's FLEX render hook never ran for an FM Synth track, so neither the frame guard nor the CPU budget (Improvement: a CPU budget admits Octatrack machines before they play #1482) counted, shed or silenced its voices. Three FM Synth tracks beside a busy eight-track pattern froze the unit on PLAY.

• What this changes: The base checks every renderer at the frame builder's two calls (0x4000d43e, 0x4000d52c, two more frame-guard sites). Stock's renderers and the base's own go straight through. A renderer a module put there itself goes through mw_foreign_render:

  • its machine's track is refused (CPU budget, START, unpaid voice) or guarded with the silent source, like any other machine;
  • a STATIC or FLEX track nobody owns is guarded with stock's own render in its place, so it keeps playing.
  • No module changes. FM Synth's sites still install beside the new ones (catalogue_overlaps passes). New beta base d377cece; the five base-bound packages are relinked to it, their code unchanged.
  • Behaviour change: in a busy project the guard can now drop FM Synth notes that played before.

• How I checked it: npm run pr:ready says Ready on main 1f5dcae41; the registry host test covers the guard, the stock fallback and the CPU refusal on the new path. On the owner's MKII, a RAM-booted dev twin of this branch (c44b117b), OS 1.40C, FM Synth 0.1.2 as a trial, one short run each:

  • cpu shows the FM Synth track (7.6 % of a frame at its longest); on the twin before, 0.0.
  • With the allowance set to 5 %, the track is refused and its render costs 0.0.
  • Three FM Synth tracks beside a busy eight-track pattern: played and stopped, frame max 124.6 %. The same setup froze the unit on the twin without this change (dfbf58d6).
  • With one FM Synth track over that pattern the guard skipped tens of thousands of renders across all eight tracks within seconds.
  • Not done: nobody listened, so how audible the dropped notes are is unknown; the release base d377cece itself was not run on a unit, only its dev twin; the trial counted one loader refusal on this twin that I did not trace (FM Synth installed and ran).

• What the guides lacked or got wrong, and where I fixed it: nothing said that a module-written renderer escapes the guard; docs/MODULE_PERFORMANCE.md now says the base checks every renderer at the two calls.

Related to #1473.

🤖 Generated with Claude Code