test(serve): align Windows schema tests with current contracts

[Apoze]

test(serve): align Windows schema tests with current contracts

Problem

ninfer_openai_schema_test no longer matched three contracts already implemented by the Windows
server branch:

• MSVC requires the speculative backend enum to be qualified with its ninfer namespace;
• conflicting enable_thinking=false and reasoning_effort=low is rejected by the request parser,
  before prompt semantics are resolved; and
• /props exposes n_ctx directly under default_generation_settings, not under params.

The stale assertions caused a compile failure followed by a failing schema test even though the
server behavior was correct.

Design and affected behavior

Update only tests/test_openai_schema.cpp to assert the current parser and /props contracts and
to use the fully qualified enum. No serving, inference, model, sampling, CUDA, or public protocol
behavior changes.

Verification

Windows 11, MSVC 19.39 / Visual Studio Build Tools 2022, CUDA 13.1, RTX 5090:

cmake --build C:\src\ninfer-build -j --config Release --target ninfer_openai_schema_test

Result: target built successfully.

ctest --test-dir C:\src\ninfer-build -C Release -R "^ninfer_openai_schema_test$" --output-on-failure

Result: 1/1 passed.

git diff --check b686696eebd43b72f58239740ae002fad6afaf5e..7ec96c7e50988e3e053d959640b46ded32700944

Result: passed with no output.

The full suite was not rerun because this is a test-only correction in one translation unit and
does not change production code.

AI disclosure

The implementation was produced with OpenAI GPT-5.6 Sol (Codex, xhigh reasoning). I reviewed the
complete diff, verified each assertion against the current implementation, and ran the focused
Windows build and test above.