feat(cpp_server): centralize errno classification + add TCP relay server

[minjcore]

Summary

Three related pieces of work on base/cpp_server:

1. os_error.h — single source of truth for errno handling. Maps raw
   errno to an app-level IoOutcome (Retry / WouldBlock / PeerGone /
   FdExhausted / NoBufferSpace / Fatal), plus an io_should_wait() helper
   for the immediate-I/O fast paths.

2. Wired classify_errno() into all three event-loop backends, replacing the
   scattered errno == EAGAIN || EWOULDBLOCK / == EINTR checks:

   • event_loop_kqueue.cpp (5 sites)
   • event_loop_epoll.cpp (5 sites)
   • event_loop_uring.cpp (4 sites — io_uring uses negated errno, so it
     classifies -ret / -res)

   Behavior is unchanged — pure de-duplication. Mapping is exact:
   EAGAIN/EWOULDBLOCK → WouldBlock, EINTR → Retry, ECONNABORTED → PeerGone.

3. tcp_replay_server.cpp — standalone TCP relay (machine A → relay →
   machine B) with bidirectional pumping and a --selftest mode that needs no
   external hosts or capture files. Added a make replay target.

4. MACOS_USERSPACE.md — notes on the macOS/kqueue userspace path:
   syscalls crossed, the kqueue event model, the errno→decision table (now
   implemented in os_error.h), macOS-vs-Linux gotchas, and running on port 443
   against ISP interference.

Test plan

• macOS: make cpp_server + make test → test_http_parser, test_tcp_ring
  pass, clean under -Wall -Wextra.
• make replay && ./tcp_replay_server --selftest → A→relay→B→relay→A round-trips
  1 MiB byte-exact, exit 0.
• git grep confirms no raw errno == EAGAIN/EWOULDBLOCK/EINTR checks remain in
  the event loops.

Notes

event_loop_epoll.cpp and event_loop_uring.cpp are Linux-only and were not
built locally (macOS). The edits are mechanical token replacements through the
shared os_error.h, which is compile-checked. Please confirm a Linux build
(CI or a Linux box) before merge.

🤖 Generated with Claude Code