perf(completion): make terminal waits lock-free

[ryux1]

Closes #246.

Summary

• Make completed and abandoned Completion waits lock-free with one OnceLock<Option<T>>; pending registration remains cancellable.
• Reuse WakerSet::register, consolidate overlapping lifecycle tests, and cover terminal reads, pending repolls/cancellation, and notification fan-out in benchmarks.

Design Notes

An uninitialized cell means pending; a stored Some(value) means completed and a stored None means abandoned. The single completer detaches the waiter set before initializing the cell, so observing a terminal result also observes waiter detachment. Pending polls and cancellation recheck the cell under the waiter mutex. Waker cloning follows the existing short-lock registration policy; replaced/cancelled wakers are dropped and notifications run after unlocking.

Local measurements compare main (f056962) with this branch on Apple M4 Max, macOS 26.6.2, Rust 1.98.0. Both use the same benchmark harness, excluding observer-container construction and handle cloning from fan-out. Values are medians of six per-run medians in nanoseconds, with run order balanced across main, an intermediate AtomicBool variant, and this implementation.

Benchmark	main	This branch	Change
ready_wait	6.117	3.107	-49.2%
abandoned_wait	6.139	2.738	-55.4%
cancel_pending	11.700	12.150	+3.8%
repoll_pending	5.280	5.492	+4.0%
complete_then_wait	65.965	62.005	-6.0%
notify_pending	94.505	90.465	-4.3%
fanout/2	174.150	164.350	-5.6%
fanout/8	420.250	387.500	-7.8%
fanout/32	1292.000	1164.000	-9.9%

Terminal reads improve substantially; pending cancellation and repolling pay roughly 0.45 ns and 0.21 ns respectively. Small fan-out differences are noisy. Completion/notification throughput is comparable to the AtomicBool variant, without maintaining a separate terminal flag.

Validation: cargo +stable x test and cargo +1.86.0 x test (484 tests each), cargo x check, cargo x lint, cargo +stable x bench --no-run, and all 14 Completion integration tests under Miri with strict provenance, symbolic alignment, and seeds 0..4.

[ryux1]

Rebased onto the current main (64d7a94) and resolved the changelog overlap, preserving the new MPSC entries alongside this PR’s completion entry. The full project gates pass on head 7010f81: cargo x check, cargo x test, cargo x lint, and cargo x miri. This is mergeable and ready for review.

[QwQBiG]

@ryux1 Thanks for your contribution.

CI is green now. I also tested 7010f81 locally on Linux (WSL): cargo x test and cargo x check passed, along with the Completion and trait tests on Rust 1.86.

Seven additional local tests covering reentrant waker callbacks, panic cleanup, and completion/cancellation races passed on stable, Rust 1.86, and under Miri with four seeds. I didn't find any blocking correctness issues. I haven't independently reproduced the benchmark comparison.

The conflict overlaps with the test-helper consolidation in #293. main now imports shared helpers from tests_integration, while this PR adds clone-callback support in the same area of completion_test.rs.

Could you preserve the shared-helper refactor and adapt the new clone-callback regression test on top? In particular, the shared assert_completes_without_deadlock now takes only the test closure. We can then recheck the updated version.

[tisonkun]

Generally LGTM.

I'm considering two minor improvements. Would test out tomorrow and get this merged.

[tisonkun]

Thanks for your contribution @ryux1 !