fix(oneshot): allow reclamation after sender handoff

[tisonkun]

Summary

Fix undefined behavior when a oneshot receiver frees the channel after the sender publishes MESSAGE or DISCONNECTED, while finish_sender_awakening still holds its &self argument. Pass the original NonNull<Channel<T>> instead, and stop using the local channel reference at publication so it does not outlive the ownership handoff.

Add coverage for concurrently dropping a waiting receiver and its sender, retain the existing completion-race regression tests, and update the changelog because the faulty implementation is present in v0.7.3.

Document the crate-rooted import convention in AGENTS.md and apply it throughout the oneshot module.

Validation:

• Reproduced the CI failure on unmodified main at 6e8762c, using nightly-2026-10-11 and the Linux x86_64 Miri target: seed 1 invalidates the protected &self argument during receiver-side Box::from_raw.
• cargo x test: 594 tests passed; cargo x lint: passed.
• Full cargo x miri: passed on nightly-2026-10-11, targeting x86_64-unknown-linux-gnu, with the original strict-provenance and symbolic-alignment flags and seeds 0..4.
• The 20-test oneshot integration suite also passed all 64 seeds in 0..64 with the same toolchain, target, and checks.

[orthur2]

The fix looks right to me. I just ran into the same Miri failure on #359 and ended up with essentially the same change.

One concern is that nothing here would catch this if it came back. With asyncband/src/oneshot reverted to main, this branch's oneshot_test still passes all 16 seeds I tried under the CI setup. The window is only between the final swap and the helper's return, and a single race rarely lands in it. A test that registers the waker first, races the sender against the receiver completing, and drops the receiver before joining the sender hits it much more often when it repeats under Miri. In my run, 128 rounds failed the reverted code at about half the seeds, including two of the four that CI runs, at about 6 seconds per seed. Would it be worth adding something along those lines here? Not blocking.

[tisonkun]

Would it be worth adding something along those lines here? Not blocking.

You can draft a PR and I'll read it out.

[orthur2]

Would it be worth adding something along those lines here? Not blocking.

You can draft a PR and I'll read it out.

Sure, opened #363.