Windows OpenAI socket smoke terminates instead of reporting a failed response

[localai-bot]

Summary

PR #524 run 31641616323 reaches the native Windows OpenAI API test, serves the final real-socket chat request, then exits with -1073740791 (0xC0000409) before doctest can print its result.

Evidence

• Windows CPU job: https://github.com/mudler/vllm.cpp/actions/runs/31641616323/job/94265239385
• test_openai_api_server.exe logs the final response from the first real-socket test, then the process terminates with no doctest summary.
• tests/vllm/entrypoints/openai/test_api_server.cpp starts a raw std::thread, has aborting REQUIRE assertions after the thread becomes joinable, and only stops/joins at the normal end of the test. A failed response therefore unwinds through a joinable std::thread and calls std::terminate, masking the assertion.
• The vendored cpp-httplib v0.49.0 closes each accepted socket immediately after writing. Upstream commit 8e702d3837b2164765ca1d98cb6d180ae4711e70 explains that this can generate an abortive RST on Windows and make a fully-written response appear as a failed read, then fixes it by half-closing and bounded-draining before close.
• Upstream commit ae8356d86eabfd3ad4a969b55266fb3ecc2aa834 independently documents the same joinable-thread assertion masking and adds scoped teardown.

Required fix

1. Make every real-socket OpenAI test tear its server thread down during assertion unwinding, so doctest failures remain observable.
2. Backport the minimal upstream accepted-socket drain fix, preserving its 100 ms / 1 MiB bounds, so Windows clients do not see a successful response as a reset.
3. Add RED-first behavioral coverage and mutation evidence for both guarantees; retain all existing request/status/body assertions and Windows release test execution.
4. Rerun native Windows CPU and Vulkan jobs, the full PR gate, and the exact merged-SHA ten-platform release dry run before tagging.