Fix input-write deadlock in UserTerminalHandler on large input bursts

[Kronuz]

The bug

The per-session terminal pump, UserTerminalHandler::runUserTerminal, is a single-threaded select loop that wrote client input to the pty master with a blocking RawSocketUtils::writeAll(). A burst of input larger than ~one pty buffer deadlocks the session:

1. The input echoes back and fills the pty's output buffer.
2. With its output buffer full, the shell stalls and stops reading input.
3. The blocking write to the master never drains, so it never returns.
4. Stuck in that write, the loop also stops draining output -- the very thing that would unstick the shell.

The whole session wedges. It's reproducible by pasting a multi-KB heredoc: ~1KB on macOS, ~8-10KB on Linux (where one pty buffer is larger).

Related is issue 682.

The fix

The read path already tolerated EAGAIN -- it was written expecting a non-blocking master -- but nothing ever made the master non-blocking. So:

• Set O_NONBLOCK where the fd is created, in UserTerminal::setup (PseudoUserTerminal and the test terminals). This is now part of that interface's contract.
• Rework the pump to never block: buffer pending input, drain it to the master through the select write set, and keep reading output every iteration.
• Add backpressure: when the input buffer fills, stop reading more from the router so backpressure reaches the client -- symmetric to the existing output gating on routerWritable.

Testing

A new regression test, TerminalTest "LargeInputNoDeadlock", drives a real pty (cat in raw mode) through the real handler. It first does a small warmup round-trip to confirm the full client -> pty -> client echo path is live, then pushes an 8KB payload (well past one pty buffer) and requires every byte to echo back within a generous deadline. The old code deadlocks and the read never completes, so the test times out and fails; with the fix it round-trips and passes. The warmup keeps the test reliable on a slow build (a sanitizer run under ctest --parallel), where it would otherwise start typing before the pipeline was wired; it is far smaller than one pty buffer, so it round-trips even on the buggy code and cannot mask the regression. The existing tests never caught this because they all use FakeUserTerminal (a socket pair), which has no pty buffer semantics.

The full et-test suite is green, and the fix has been validated live: a 56KB one-shot input that previously wedged the session now round-trips byte-exact against a real etserver.

Why it's low-risk

Server-side only -- no protocol, client, or proto changes. The change is confined to the terminal pump and the UserTerminal::setup contract; the wire format and everything around it are untouched. A non-blocking master is what the read path already assumed, so this makes the code match its own expectations.

Asking for guidance

Setting O_NONBLOCK in setup (rather than inside the pump) made the contract explicit and kept the fake terminals honest, but I'm happy to move it if you'd rather. Is the approach right?

---

This PR was prepared with AI assistance; I've reviewed and stand behind it.

[codecov]

Codecov Report

❌ Patch coverage is 86.55462% with 16 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.06%. Comparing base (6368584) to head (12c4a90).

Files with missing lines	Patch %	Lines
src/terminal/UserTerminalHandler.cpp	59.09%	9 Missing ⚠️
test/integration_tests/TerminalTest.cpp	92.55%	7 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #765      +/-   ##
==========================================
+ Coverage   86.70%   87.06%   +0.36%     
==========================================
  Files          72       72              
  Lines        5460     5574     +114     
  Branches      508      525      +17     
==========================================
+ Hits         4734     4853     +119     
+ Misses        726      721       -5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.