[iamwavecut]

Fixes the actionable part of #114: on Qwen3.8-Flash-Next (and every Qwen preset) a request that offers tools often got one tool call where two or three were needed.

Cause. The non-thinking preset's presence_penalty 1.5 (the model card's recommendation) subtracts 1.5 from the logit of every token the answer already used. A second tool call repeats <tool_call>, <function=, the function name and <parameter=, so near-greedy sampling turned the second call into the end of the turn or prose. The reporter measured BFCL parallel_multiple 59.5 with the preset and 67.0 with --presence-penalty 0 (llama.cpp, which applies no penalty, 70.5).

Changes

• src/serve/translate.cpp: a request with tools and a tool_choice other than none resolves presence_penalty to 0 unless the request or --presence-penalty names a value; the post-thinking sampler follows the same rule. All three protocols go through this point. No tools, or tool_choice: none, keeps the preset.
• src/serve/translate.cpp, src/models/qwen3_5/frontend/chat_template.cpp: a tool definition is rendered as {"type":"function","function":{"name","description","parameters"}} in that order, the order the Qwen templates were trained on; strict stays in the tool JSON for the output contract (--structured-output grammar) and the frontend strips it from what the template renders. The reporter found NInfer rendering name, parameters, strict, description; the effect of that on the score was not isolated.
• Tests: test_openai_schema covers the rendered member order and the penalty rule (tools → 0, no tools → preset, tool_choice: none → preset, request value and --presence-penalty kept, post-thinking preset and request value); test_frontend renders a strict tool through the Qwen3.6 and Qwen3.8 templates and checks the order, the absence of strict and that the contract still sees strict.
• Docs: serving (function tools, sampling presets), cli (preset table pointer).

Verification. ninfer_qwen3_5_frontend_test built and run on the Mac through a scratch host-only CMake (passes; fails against the old renderer). translate.cpp and test_openai_schema.cpp pass clang -fsyntax-only; the serve tests link the engine and were not run here. CTest, CUDA builds and CI were not run locally. BFCL was not re-run: the expected effect is the reporter's --presence-penalty 0 result (59.5 → 67.0 on parallel_multiple, irrelevance unchanged); the remaining gap to llama.cpp (~3.5 points, two cases answering in prose) is not addressed here.

CI. The MSVC host-compile checks were red on master before this PR (since 86dd954: _cvtsh_ss in src/ops/sparse_moe/streamed_cpu_avx2.cpp and M_PI in tests/ops/test_gguf_expert_cpu.cpp). The third commit fixes both with the vector F16C form the GGUF expert kernel already uses and std::numbers::pi; both files pass a clang x86-64 cross syntax check with AVX2/FMA/F16C locally, MSVC itself runs only in CI.