better async model

[ceifa]

Why

The async model had one primitive — yield the Lua coroutine and let a JS loop sniff the last yielded value — and everything (awaiting, time-slicing, cancellation, callbacks) was squeezed through it. That is why :await() failed inside table.sort/gsub/promise:next callbacks and in JS→Lua callbacks, why top-level coroutine.yield values were dropped, why a parked run couldn't be timed out or aborted, and why a loop of awaits starved the event loop.

Full design write-up and the measurements behind it: docs/async-redesign/README.md.

What changed

Two engines, chosen automatically at load (async: 'auto' | 'jspi' | 'yield').

• JSPI (default where available — Chrome/Edge 137+, Firefox 153+, Node 25+). The interpreter runs under WebAssembly.promising, so an :await() suspends the whole wasm stack and works anywhere: inside a table.sort comparator, a gsub callback, a promise:next handler, a coroutine nothing drives from the host, or a Lua function called back from JS.
• Coroutine yielding (fallback). Same observable behaviour everywhere it can express it; it just can't cross a C-call boundary, which now raises a clear error instead of a misleading one.

Highlights

• :await() across C-call boundaries under JSPI (previously an error).
• JS→Lua callbacks are sync-or-promise: a callback that awaits returns a Promise, one that doesn't stays synchronous with no promise allocated.
• A run parked on a promise is interrupted promptly by timeout / AbortSignal instead of waiting for the promise to settle; the interrupt is not swallowable by a script pcall.
• A top-level coroutine.yield that isn't an await is a host yield: new run(..., { onYield }) receives its values and decides the resume values (previously silently dropped); an unrepresentable yielded value no longer crashes the run.
• Resumes go through a MessageChannel macrotask, so timer-driven code no longer starves and browsers avoid the 4 ms nested-setTimeout clamp.
• Closing a state while a JSPI run is parked rejects the run rather than resuming into freed memory.

Build

• glue.wasm is now built with -sSUPPORT_LONGJMP=wasm (required for JSPI; also ~30–45% faster on pcall/error/yield paths, size-neutral). src/native/wasmoon.c gains a small C trampoline that reaches the suspending import only from a wasm frame, plus the existing lua_Integer double-twins.

Cost / benefit (measured, Node 26)

• Interop and CPU paths unchanged (heapsort, Call Lua from JS, Call JS from Lua flat; lua_resume ≈ lua_pcallk per callback).
• An await of a settled promise drops from ~6 µs to ~0.3–1.6 µs under JSPI.
• 1000 concurrently parked runs cost ~4 MB, same as before (per-suspension C-stack copy, not fixed regions).