Add cede() for yielding the current virtual thread

[adamw]

Adds cede() to ox.control, next to checkInterrupt():

inline def cede(): Unit =
  checkInterrupt()
  Thread.`yield`()

Virtual threads are not preempted, so a CPU-bound loop can starve other virtual threads (in an experiment with 2x-cores busy-looping VTs, half of them received zero CPU time). cede() gives compute-intensive code a way to cooperate with the scheduler — and, since it also checks the interrupt flag, with the cancellation protocol: a busy loop calling cede() becomes both fair and cancellable.

Why checkInterrupt + Thread.yield

Based on experiments comparing yielding techniques on virtual threads (JDK 21 & 26, Linux, 8 cores), following up on this gist:

• Thread.yield genuinely unmounts and resubmits the continuation to the back of the scheduler's queue (external submit when the carrier's local queue is empty) — it reliably prevents starvation (89–93% liveness-probe capture vs 2% baseline), at ~1µs per call.
• LockSupport.parkNanos(1) gives near-perfect round-robin fairness, but costs ~40–46µs per call (timer round-trip), stalls the caller for that time, and silently degrades to a no-op when the interrupt flag is set — i.e. exactly when the enclosing scope is being cancelled. Since Thread.yield is ~50x cheaper, calling it more frequently (every ~1ms of computation) matches parkNanos(1)'s liveness at a quarter of the overhead. The trade-off is mentioned in the scaladoc.
• LockSupport.parkNanos(0) is a true no-op.
• Thread.sleep(0) behaves identically to checkInterrupt + Thread.yield (that's literally its JDK implementation for virtual threads), but only as an @implNote — the explicit version doesn't rely on unspecified behaviour.

The interrupt check makes cede() a cancellation boundary, consistent with e.g. cats-effect's IO.cede.

Includes tests and a docs entry in utils/control-flow.md.

🤖 Generated with Claude Code