Add Codex Responses prompt normalization

[krystophny]

Summary

• normalize OpenAI Responses developer content into a single leading system message
• merge instructions with developer/system content in the shape Codex expects
• add targeted regression tests for Codex prompt compatibility on top of the Responses API core

Stack position

• depends on krystophny/vllm-mlx#1: Add OpenAI Responses API core #1
• intentionally does not change the core /v1/responses transport or persistence logic
• this is the client-compat layer needed for Codex local mode after the core endpoint exists

Why this is independently deployable on top of #1

• only affects /v1/responses
• leaves chat/completions and Anthropic message paths alone
• makes the Codex-specific prompt normalization explicit instead of smuggling it into the core endpoint PR

Related upstream context

ggml-org/llama.cpp

This PR follows the same class of fix that already showed up in llama.cpp:

• llama.cpp#20079 translated developer -> system and merged prompt pieces for Codex/template compatibility: fix: translate "developer" role to "system" for chat template compatibility ggml-org/llama.cpp#20079
• llama.cpp#18486 and #19720 provide the broader Responses background there: server: /v1/responses (partial) ggml-org/llama.cpp#18486 server: add OpenAI Responses API compliance ggml-org/llama.cpp#19720

vllm-project/vllm

Related open Responses prompt/state fixes upstream include:

• vllm#37727 instructions leaking with previous_response_id: [Bugfix] Fix Responses API instructions leaking through previous_response_id vllm-project/vllm#37727
• vllm#37739 default chat-template kwargs handling in Responses: [Frontend] Fix default_chat_template_kwargs handling in Responses API vllm-project/vllm#37739

Our scope here is narrower: normalize the prompt shape so Codex local turns render cleanly against vllm-mlx.

Validation

• PYTHONPATH=/Users/ert/code/vllm-mlx /Users/ert/code/.venv/bin/python -m pytest tests/test_responses_api.py -q
• local Codex validation against vllm-mlx in the FortBench MLX rerun

What could still improve

• additional tests covering more Codex prompt combinations with previous_response_id
• explicit fixtures for unsupported built-in tools alongside Codex prompt normalization
• upstreaming once the core Responses PR shape settles