Verdict on this Mac: PASS. Linux: PENDING until the workflow's ExUnit run includes test/sched_test.exs.
abi_dirty_cpu() and abi_dirty_io() are generated from ERL_NIF_DIRTY_JOB_CPU_BOUND and ERL_NIF_DIRTY_JOB_IO_BOUND. scripts/check-abi.sh diffs those integers with the header. On this OTP they are 1 and 2.
Mojoler.Sched.burn/1 and hold/2 are registered with the CPU flag. io_ping/0 uses the I/O flag. The DSL generator rejects any other schedule name.
E22: in a VM started with +S 1, a 200 ms dirty burn let 18 heartbeats through. The same loop with flags 0 let 0 through. The normal-scheduler library is experiments/sched/normal_burn.mojo, built into a temp directory.
hold/2 copies the binary into a calloc buffer, frees the inspect descriptor, then spins, then writes a new binary. The borrowed view does not span the spin.
Task.yield(task, 30) returned nil while burn(200) was running. Task.await then returned :ok. The timeout stopped the wait, not the native loop. There is no cancellation API.
A dirty scheduler keeps the normal scheduler free. It does not contain a crash. The abort probe is a separate process for that reason.