Reproducing a global memory budget stall needs pinned head replies

To show that a server's small replies stall behind a global reply budget that other clients fill, unread replies are not enough: small replies drain into kernel socket buffers and get released. Pin each load connection with a big head reply first (several MiB, bigger than the socket buffers, relay with a small SO_RCVBUF that never reads), so every later reply on that connection queues behind an unsent segment. Then fill the remainder with descending power-of-two reads, one size per connection, so per-connection in-order blocking doesn't stop the fill at one level. Finish with a probe read larger than the per-connection small allowance: if it still waits, the budget really is full. Watch for the server's idle-reply timeout, which caps how long the stall can last.