Refactor freelist MPSC queue draining

[Matthew Parkinson (mjp41)]

Summary

• make FreeListMPSCQ::enqueue() report the empty-to-nonempty transition
• add concurrent drain_and_reset() and use it consistently from allocator flush and queue cleanup
• share bounded chain traversal between ordinary dequeue and drain while preserving ordinary dequeue's retained final tail
• remove redundant queue initialization and quiescent destruction APIs
• add focused sequential, publication-gap, callback-stop, concurrent producer, storage-reuse, and callable-compatibility tests

Queue semantics

The representation remains message-pointer based: an empty queue has null front and back, and a non-empty queue has front at the first message and back at the final message. Ordinary dequeue still retains the final message.

drain_and_reset() clears and detaches one queue snapshot before invoking callbacks. Producers whose back exchange observes the reset form a replacement queue that is left for a later drain. The operation waits indefinitely for a producer that has exchanged back but has not yet published front or its predecessor link; this preserves the previously agreed fork/interrupted-producer tradeoff.

The shared process_chain() primitive processes objects strictly before a fixed address bound and returns the first unprocessed object. Dequeue retains that object; drain retries an unpublished link until it reaches and consumes the captured target.

Validation

• final Debug suite: 83/84 on the initial run and both required reruns; only the known perf-lotsofthreads-check 60-second timeout remains
• func-malloc-fast and func-jemalloc-fast: pass
• ThreadSanitizer heavy queue stress: 20/20 pass
• forward-edge and backward-edge focused configurations: 4/4 each
• available mitigation executables: 5/5 pass; 47 registered variants remain unavailable in the existing local mitigation build configuration
• exact-name Release codegen comparison: all 12 dequeue instantiations retain their prefetch counts and are slightly smaller; no out-of-line process_chain() symbol or call
• powered Release comparison: perf-msgpass-fast --smoke median 0.88s to 0.64s; perf-lotsofthreads-fast --smoke median 19.945s to 17.955s
• final fresh-context implementation review: approved

The repository-provided clang-format 15 binary is x86-64 and cannot run on this ARM64 host; the five changed source files were formatted and verified with clang-format 16, and git diff --check passes.

Deferred work

This PR does not implement allocator sleeping or retirement, pool scavenging or other pool policy, retained-tail retirement, wake counters or hints, or the N+1 benchmark. Those remain separate allocator-lifecycle work.

[Nathaniel Wesley Filardo (nwf)]

Hm. It's been a while, but I thought the two of these were there because pointers to freelist heads and intra-list link pointers were different types?

[Matthew Parkinson (mjp41)]

I think you're right. I think this and the original were incorrectly domesticating the head with a link domesticator.

[Matthew Parkinson (mjp41)]

Fixed in d7f262e. drain_and_reset now takes distinct head and queue domesticators, matching dequeue. CoreAllocator::flush uses the trusted annotating conversion when QueueHeadsAreTame and validates the head otherwise; decoded message links always use the queue domesticator. Added focused separation, rejected-head, and rejected-successor tests.