observe() is called once per emulated instruction, and nearly every one of
those is a step forward inside the same code — not a backward branch, not a
long jump. It was doing the loop-close bookkeeping first anyway: two range
updates and a counter on fields of a foreign object, then the test that would
have told it none of that was needed yet.
The forward case now returns after one comparison and at most one field write.
The loop's lowest PC is only updated where it can actually change — a backward
branch IS the loop top — so nothing about what counts as an idle spin moves.
Measured on the RP2350 host (BadgeOS, MicroPython on the M33), together with a
one-read-per-turn change on the caller's side: +11% instructions per CPU
second, four rounds out of four.
b2f5a4f1f3