Pool worker fibers and coalesce reactor wake-ups

[himura467]

Summary

Two fixed-cost reductions motivated by the boundary benchmark (#11):

• Fiber pooling: after finishing a request, a worker fiber parks itself back into the pool and is resumed with the next job, instead of being created and GC'd per request. Pushing before yielding is safe because the single-threaded VM cannot pop the fiber until the yield suspends it; dead fibers are skipped on pop.
• Wake coalescing: nsgi_handle only writes a wake byte when it flips the wake flag false to true; the reactor resets the flag when a pop finds the queue empty. Both transitions happen under the queue mutex, so a producer that skips the write is guaranteed its job is seen by the drain loop the in-flight wake triggers (no lost-wakeup window).

Benchmark impact (Apple M2, 12-header baseline)

Metric	Before	After
8-thread throughput	161k req/s (6.2 us/req)	336k req/s (3.0 us/req)
Single-threaded	60k req/s (16.5 us/req)	99k req/s (10.1 us/req)
Allocations per request	12.0	10.0
Crossover vs idealized Rack	~60 headers	~22 headers

Scaling stays flat in header count and body size (60 headers: 339k req/s; 16 KB body: 338k req/s; 10 allocations in every scenario), which puts the crossover against the handoff-free idealized Rack harness inside the 15-25 header range of ordinary browser traffic. bench/README.md updated with the new matrix.

Verification

• Mock host suite: 239 requests, all paths green, interleaving at 212ms.
• leaks --atExit: 0 leaks.

🤖 Generated with Claude Code