fix(xz): deterministic async XZ decoder wakeups, no worker deadlock [DEV-124]

[P4suta]

Fixes the flaky async XZ decoder hang that intermittently timed out CI (big-endian (s390x, qemu), test (macos-latest), test (ubuntu-latest)), blocking the RM-400 all-CI-green gate. Linear: DEV-124.

Root cause

The portable XZ decoder (filter/xz.rs) drives a blocking pull decoder (lzma_rust2::XzReader) on an OS worker thread over two sync_channel(2) channels. Codec::process mixes non-blocking poll_event (try_recv) with a blocking wait_event (events.recv()). Because the async adapter AsyncCodecReader::poll_read calls process inline, wait_event blocks the executor thread. Once the NeedInput event has been drained by an earlier try_recv, the worker waits for input while the caller waits for an output event with nothing in flight — a mutual wait that slow qemu/loaded-macOS scheduling turns into a ~14-minute hang. Native (liblzma) builds are unaffected.

Fix — worker→task waker + non-blocking codec drive

• core Codec gains a defaulted poll_process(.., waker) (non-worker codecs inherit it; only XZ overrides).
• XZ gains an EventSink { SyncSender, Arc<Mutex<Option<Waker>>> } that wakes the registered task waker after every event send (both NeedInput and Output). process is refactored into drive(waker: Option<&Waker>): None keeps the byte-for-byte blocking sync path; Some(w) registers the waker (via Waker::will_wake) before the deciding try_recv and returns Ok(None) → Poll::Pending on Empty.
• Completion race: decode_worker runs the decode loop in an inner scope so all event senders drop (channel disconnects) before it wakes via a retained waker Arc — so a parked owner's next try_recv always observes Disconnected, never a lost Empty.
• pipeline_codec.rs mirrors poll_process; async_filter.rs::poll_read calls poll_process(.., cx.waker()) and maps Ok(None) → Poll::Pending.

No unsafe, static dispatch preserved, bounded memory unchanged, sync Pipeline behavior byte-for-byte identical, Sync/RefUnwindSafe preserved. Not a skip — the deterministic wakeup is the fix.

Tests

New async_xz_never_deadlocks_under_slow_source drives the async XZ reader over a 1-byte-per-poll source × N=50 under a park/unpark executor and a process-abort watchdog (a hang fails fast instead of eating the 15-min budget); the three previously at-risk tests are routed through the same harness. async_stream_v2 passes on portable and native profiles.

🤖 Generated with Claude Code