Wire unchanged: mixes freely with v3.2–v3.5. No deploy strictly needed over v3.5.0 (same wire, same behavior on the fast path).
- Pooled CSPRNG: one 64KB kernel read sliced per record instead of 1–2 getrandom syscalls each — bulk throughput up ~10–20%
- Deadline discipline: writers are write-only (SetWriteDeadline), so a future short read deadline can never hang a sibling pump; documented in code
- Debug record counters (recs/payload per session/trunk) for performance work
- Reverted an experimental 2ms send-batcher: measured zero bulk gain (reads arrive full) against a proven +4ms trickle tax