Canonicalize volatile system prompt headers in OpenAI paths

[Thump604]

Refs #524.

Summary

• Add a small static system-prompt canonicalization helper with the currently validated x-anthropic-billing-header stripper.
• Apply it to prepared system-role messages in Chat Completions and Responses paths before engine execution.
• Cover the helper, Chat Completions preparation, Responses preparation, and existing Anthropic adapter behavior.

Local repro / observed behavior

On current waybarrios/vllm-mlx main (f068991 when this branch was cut), the Anthropic Messages adapter removes x-anthropic-billing-header: from request.system in vllm_mlx/api/anthropic_adapter.py::anthropic_to_openai, but the OpenAI server paths did not apply the same canonicalization:

• vllm_mlx/server.py::_prepare_chat_messages
• vllm_mlx/server.py::_prepare_responses_request

A Chat Completions system message or Responses instructions value containing:

x-anthropic-billing-header: account=abc; cch=rotating-hash

was still present in the prepared system message sent to the engine.

Expected behavior

The validated billing-header line is non-semantic request metadata and should be removed from system-role text before engine execution. User-role content with the same text is preserved, and user-visible timestamp text is not stripped.

Minimal patch shape

• New vllm_mlx/api/prompt_canonicalize.py module with a static stripper list and canonicalize_system_prompt().
• New canonicalize_system_messages() helper that copies only changed system messages.
• Call the helper after existing message normalization in Chat Completions and after Responses input conversion.
• Do not add runtime registration APIs or speculative strippers.

Explicitly not claimed

• This does not add timestamp, MCP UUID, or session-ID strippers.
• This does not change SimpleEngine system-prefix KV-cache logic from feat: extend system-prompt KV cache to pure-LLM stream_chat path #523.
• This does not include new TTFT or cache-hit-rate benchmark results.
• This does not change media extraction, tool parsing, sampling, or decode controls.

Verification

AI_RUNTIME_BYPASS_SAFETY_GATE=1 PYTHONPATH=/opt/ai-runtime/worktrees/vllm-mlx/issue-524-prompt-canonicalization /opt/ai-runtime/venv-live/bin/python -m pytest tests/test_prompt_canonicalize.py tests/test_responses_api.py tests/test_anthropic_adapter.py tests/test_server.py::TestPromptCanonicalization -q
# 74 passed

uvx ruff check vllm_mlx/api/prompt_canonicalize.py vllm_mlx/server.py tests/test_prompt_canonicalize.py tests/test_server.py tests/test_responses_api.py
# All checks passed

/opt/ai-runtime/venv-live/bin/python -m black --check --target-version py312 vllm_mlx/api/prompt_canonicalize.py vllm_mlx/server.py tests/test_prompt_canonicalize.py tests/test_server.py tests/test_responses_api.py
# 5 files would be left unchanged

git diff --check
# clean