mempipe: make the crate testable under Miri

[gamozolabs]

Follow-up to #22: the Miri harness used to verify that fix, plus the reproducer example mentioned there.

Miri support (4e0237c)

Miri cannot execute shm_open or mmap, so this adds a heap-backed pipe that exists on every target: SendPipe::create_local() and the unsafe RecvPipe::open_raw(). The sushi_roll paths become thin wrappers over them and pipe init is shared by both constructors. No cfg(miri) in library code; the only Miri-specific lines are #[cfg_attr(miri, ignore)] on the three shm-based tests and the size constants in the new test.

It also fixes a latent provenance bug: both chunk pointers were derived from chunks[ii].0[0], which under Stacked Borrows only has provenance for that one byte, so every access past it was UB. Tree Borrows accepts it, which is why it went unnoticed. Now taken from the whole array with .cast::<UnsafeCell<u8>>().

A new multi_reader test runs on a local pipe everywhere and shrinks itself under Miri so it can be swept over seeds:

MIRIFLAGS="-Zmiri-preemption-rate=0.5 -Zmiri-seed=$seed" \
    cargo +nightly-2024-03-01 miri test -p mempipe -- multi_reader

Run with and without -Zmiri-tree-borrows; neither aliasing model contains the other. On the code before #22 this fails about a third of seeds with a data race between the sender's chunk write and a reader's chunk read. On current main it is clean for 30 seeds under each model, and also under -Zmiri-strict-provenance and -Zmiri-symbolic-alignment-check.

Reproducer example

examples/hang.rs is a native reproducer for the pre-#22 bug. Random preemption never hit the race on a 16-core box because it needs two stalls at once, so this injects them: signal handlers that busy-wait a few microseconds, fired at the sender and reader threads at a bounded duty cycle. The library runs unmodified. Settings that hang the old code within seconds are in its doc comment, as are settings that hit the ticket-0 hole (sequence checked first with client_seq still initialised to 0). Miri is the better regression test; this is for watching it happen natively.

🤖 Generated with Claude Code