[Avicennasis]

Add allow_inf_nan=False to temperature and top_p on both OpenAI request models, matching the existing penalty fields. Raw NaN, Infinity, -Infinity, and overflowing 1e9999 now produce a field-named 422 before generation. Omitted/null fields and finite values keep their previous behavior.

The four field constraints alone produced 500 responses: FastAPI includes the rejected float in its validation errors, and Starlette refuses to serialize it. The OpenAI validation handler keeps the 422 detail envelope and renders non-finite diagnostic values as strings. Anthropic route registration delegates non-Anthropic validation errors to that handler; its own 400 envelope is unchanged.

Idea adapted from JustVugg/colibri #2342, commit 3a329791460bdb0efd88d7a56f094e2951faa0b9, and #2370, commit ef410a3d80a0e98f465528b5842faac995c44fdc. Both projects use Apache-2.0. No donor code was copied.

Verification used Python 3.12.3 and torch 2.14.1+cpu on Linux x86_64, without a GPU or checkpoint. Some declared runtime dependencies were missing from the test environment and were installed locally; no dependency declarations were changed. From the repository root:

export PYTHONPATH=python
python -m pytest tests/server/test_openai_api.py -k sampling_routes -q --tb=short
python -m pytest tests/server -q --tb=short
python -m pytest tests -q --tb=short

The focused command reproduced 16 failures (HTTP 200), then 32 failures (HTTP 500 with the field-only fix, including combined route registration), then passed all 34 cases with the complete fix; 37 existing cases were deselected. The server suite passed all 690 tests. A local compatibility probe (not part of this change) passed 62 checks on both the base and the branch, covering omitted/null values, finite values outside the usual sampling ranges, and unchanged penalty bounds/non-finite rejection.

Full suite: branch 20 failed / 1699 passed / 919 skipped / 1 warning; base 20 failed / 1665 passed / 919 skipped / 1 warning. The exit code is 1 for both. All 2604 existing test outcomes match, the 20 failing IDs are the same known CPU-only environment failures (CUDA/flashinfer/native-extension tests) on both trees, and all 34 added cases pass. With an unwritable default TVM-FFI cache, both trees show 165 additional failures; pointing TVM_FFI_CACHE_DIR at a writable directory removes them.

NOT VERIFIED: CUDA sampling, real-model generation, checkpoint-provided non-finite defaults, direct engine callers, or non-finite sampling validation on Anthropic/Responses requests. The optional guard in resolve_sampling.pick() is outside this fix. No performance claim is made.

Based directly on 628b88ff795b5ef2c7f9baff088f0e7720d8ebb4. The open PRs #334, #459, #565, #592, #593 and draft #656 share no files with this change; no stacking dependency is needed. No maintainer decision is needed to review the fix.