[caixuf]

Root cause

After #22, persistent localhost echo is healthy (WSL ~43.7k req/s, fail=0). The remaining cost called out there still stood: every Socket::read / write / accept / connect wait did one-shot add_fd + remove_fd (EPOLL_CTL_ADD/DEL, replace handlers_[fd]).

That is extra syscalls on the steady-state path. It is also the shape of the #22 race: if the EventLoop thread set_values (worker may add_fd the next wait) and then remove_fd, the new handler is erased and the waiter hangs.

Short-connection connect+echo p99 under wave=32 is a different bottleneck: one accept() per epoll event + listen-queue physics. ADD/DEL on the listen fd made that worse, but draining the backlog cannot make the tail of a 32-way handshake wave look like p50.

Design

Keep the fd in the poll set for the life of the Socket. Close still remove_fd (DEL + erase).

1. add_fd: first wait EPOLL_CTL_ADD (store handler before ctl so EPOLLONESHOT cannot fire with no waiter). Later waits update the existing IoEventHandler in place and EPOLL_CTL_MOD. No DEL.
2. Linux EPOLLONESHOT: kernel disables the fd after one event, so a registered fd with no waiter does not LT-spin. Re-arm is the next add_fd/MOD.
3. complete_io before set_value / set_exception (the fix(net): persistent localhost echo no longer times out #22 ordering): clear waiter callbacks and interest bits, do not DEL/erase. A CoroutinePool worker that resumes before the callback returns can add_fd a new waiter on the same handler slot.
4. Windows WSAPoll: same protocol. complete_io sets events=0 and skips those fds when rebuilding the poll set (ONESHOT equivalent). Still compiles; no epoll.
5. Accept drain: on listen readability, accept4 until EAGAIN into pending_accepts_. Later co_await accept() pops without another epoll wait. Default listen(SOMAXCONN).

Not done (on purpose): no CoroutinePool/Task rewrite; no timeout bumps; no fake p99 by shrinking wave.

Tests

• Existing persistent echo (4 clients, 300ms, 0 get() timeouts) — fix(net): persistent localhost echo no longer times out #22 + re-arm.
• Existing nested Task cross-thread continuation (test_core).
• New real_net_accept_burst: 16 concurrent connect+echo, 0 timeouts.

How to verify (Linux; WSL is the performance source of truth)

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --target real_net_benchmark test_real_net_echo test_core -j$(nproc)
./build/tests/test_real_net_echo
./build/tests/test_core
./build/benchmarks/real_net_benchmark
# or:
./build/benchmarks/real_net_benchmark echo --duration-ms 1500 --clients 8 --payload 64
./build/benchmarks/real_net_benchmark connect --connections 200 --payload 64

Success:

• Persistent echo: wall ≈ requested duration, fail=0, no Task::get timed out. Throughput on WSL is the number to quote — not a cloud VM.
• Connect+echo: fail=0. p99 under wave=32 will still show accept-queue delay; that is expected.

Do not treat cloud QPS as success. Loopback only; not NIC/wire; not the CPU-sim Echo rows in professional_flowcoro_benchmark.