Finished fiber stacks went back to a spare pool that the event loop
unmapped whenever it was about to sleep. A server's loop sleeps between
nearly every pair of requests, so each request that awaited mapped, faulted
in and unmapped a fresh 512 KiB stack: on a Hono server mmap, munmap and the
page faults were about a fifth of its CPU time. Spares now outlive short
sleeps. At most once a second a sleep unmaps the spares that no fiber took
since the previous trim, and a sleep longer than two seconds unmaps them
all, so an idle program still returns the memory.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
