refactor(server): stop branching on errno after accept/recv/send

[vjan-nie]

The 42 subject forbids using errno to drive control flow after the non-blocking I/O syscalls. acceptClient/handleClientInput/ handleClientOutput each inspected errno (EAGAIN/EWOULDBLOCK) to decide whether to log or disconnect. Removed:

• accept() < 0 -> return (silent; no errno, no log)
• recv() <= 0 -> disconnect only on == 0 (orderly close)
• send() < 0 -> return
  Real socket errors (RST/ECONNRESET) are handled by the EPOLLERR|EPOLLHUP
  branch in run(); as a bonus this removes a latent false-positive
  disconnect on EINTR.

Characterization test added (RobustnessTest.AbruptDisconnectViaRST): forces a real RST via SO_LINGER and proves in black box that the nick is released (client torn down). Green before and after the refactor.

Note: which branch tears down an RST client is timing-dependent before this change (EPOLLIN+errno may win the race over EPOLLERR|HUP). After this change EPOLLERR|HUP handles it consistently, same iteration. Do not assume EPOLLERR|HUP alone covered RST prior to this commit.

Suite: 139/139.

[vjan-nie]

Closes #4

---

Problem

The event loop used errno (EAGAIN/EWOULDBLOCK) after the non-blocking I/O
syscalls to decide whether to log or disconnect. The 42 subject forbids
errno-driven control flow after accept/recv/send (eliminatory), and in
recv the check also mis-fired on EINTR, disconnecting a healthy client.

Change

• acceptClient: accept() < 0 → return (silent; no errno, no log).
• handleClientInput: recv() <= 0 → disconnect only on == 0 (orderly
  close); < 0 → return (nothing ready now).
• handleClientOutput: send() < 0 → return.

Real socket errors (RST/ECONNRESET) are reaped by the existing
EPOLLERR|EPOLLHUP branch in run(). Removing the recv check also cures the
latent EINTR false-positive disconnect.

Two files only: src/Server.cpp and the new characterization test in
tests/test_robustness.cpp. include/Server.hpp unchanged. The un-polled
send() in disconnectClient is a separate correction-sheet item and is
intentionally left for a follow-up.

Verification

• RobustnessTest.AbruptDisconnectViaRST (new): forces a real RST via
  SO_LINGER{1,0} and proves in black box that the nick is released (i.e. the
  reset client was torn down server-side), with a bounded retry (no fixed
  sleep). Green before and after the refactor — a valid characterization
  test for a refactor.
• grep errno src/Server.cpp: no errno remains in the accept/recv/send
  branches (only socket-setup throw messages, the EINTR check on
  epoll_wait itself, and epoll_ctl logging).
• make mandatory and full build: clean, -Wall -Wextra -Werror -std=c++98.
• Suite: 139/139 (138 prior + 1 new).

Note for future event-loop work (please keep in the history)

Which branch reaps an RST client is timing-dependent before this change:
when EPOLLIN and EPOLLHUP arrive in the same event, the EPOLLIN path
(recv + errno) can win the race over EPOLLERR|EPOLLHUP. After this change,
EPOLLERR|EPOLLHUP handles it consistently, in the same iteration (observable
as the RST teardown reason changing from "Connection closed" to
"Connection error"). Do not assume EPOLLERR|HUP alone covered RST prior
to this commit — the earlier audit premise was corrected during review.

What I'd like you to validate

• Builds and full suite pass in your environment.
• AbruptDisconnectViaRST is green with src/Server.cpp reverted to
  main (characterization), and green with the refactor.
• No errno left in the accept/recv/send branches.
• Scope: only src/Server.cpp + tests/test_robustness.cpp;
  disconnectClient untouched (deferred).