256 KiB fiber stack allocated for every async function

[olup]

My toy server exhausts memory very quickly because every async function call, whether awaited or not, creates a fiber with a 256 KiB stack.

This part may still be a work in progress, but a stackless coroutine implementation based on state machines would likely be much more memory-efficient.

[olup]

A prototype explored a hybrid approach: stackless state machines where supported, with smaller guarded fibers as fallback.

It was tested using a small HTTP server built specifically for the experiment. Each benchmark ran two bursts of 10,000 /health requests at concurrency 100. Results are medians from three trials on Apple Silicon macOS.

Implementation	Peak RSS	Post-load RSS	Throughput
Current 256 KiB fibers	260.98 MiB	260.98 MiB	9,627 req/s
Guarded 64 KiB fibers	32.81 MiB	4.94 MiB	12,313 req/s
Hybrid stackless + fallback	16.33 MiB	4.47 MiB	11,785 req/s

Most of the gain comes from improved fiber-stack allocation and reclamation. Stackless lowering then halves peak RSS again, with a 4.3% throughput cost versus the optimized fiber version.

The prototype supports immediate async functions, tail-await adoption, and linear awaits with heap-backed continuation states. More complex suspending control flow still falls back to fibers.

The implementation covers both C and LLVM backends and is split into incremental commits. Code and benchmark details can be shared if useful.