MemoryTransport wakes waiters with array_shift, making FIFO delivery O(n)

[nficano]

Summary

Second-pass performance/code-quality audit finding. MemoryTransport uses a PHP array as a FIFO waiter queue, then wakes the oldest waiter with array_shift, which reindexes the entire array.

Evidence

• src/Transport/MemoryTransport.php:26 declares $waiters as a list<DeferredFuture<Envelope|null>>.
• src/Transport/MemoryTransport.php:63 checks whether waiters exist.
• src/Transport/MemoryTransport.php:64 calls array_shift($this->waiters) for FIFO delivery.
• src/Transport/MemoryTransport.php:87 through src/Transport/MemoryTransport.php:95 rebuilds the full waiter list when a receiver is cancelled.

Why it matters

Every delivery to a waiting receiver is O(n) in the number of waiters. Cancellation also scans and rebuilds the entire list. The in-memory transport is used heavily by tests and in-process scenarios, so waiter-heavy tests can degrade quadratically and hide transport behavior behind queue overhead.

Proposed fix

Store waiters in SplQueue or an indexed deque-like structure. If arbitrary cancellation removal is required, assign waiter IDs and lazily skip cancelled waiters when dequeuing.

Acceptance criteria

• FIFO wakeup is O(1) in the common case.
• Cancellation does not rebuild the whole waiter list for every cancelled receive.
• Existing memory-transport ordering tests still pass, with a new test covering many pending receivers.