server: close out the upstream /v1/responses merge plan

[krystophny]

Upstream reference: waybarrios#214

Problem:
We want an OpenAI-compatible /v1/responses endpoint for local coding-agent workflows, including Codex-style request normalization. The core PR is broad and important, but it is not yet clearly prioritized against the engine/scheduler correctness backlog.

Status:

• prompt normalization from old PR responses: normalize developer and instructions for Codex waybarrios/vllm-mlx#219 is already folded into server: add OpenAI-compatible /v1/responses endpoint waybarrios/vllm-mlx#214
• no substantive review feedback is present yet
• this is the largest and riskiest open PR in the stack

Path to completion:

1. Rebase on current upstream main and confirm the diff is still coherent after folding the Codex normalization work into the main PR.
2. Re-run tests/test_responses_api.py and any broader server tests.
3. Audit unsupported semantics again so the endpoint fails explicitly rather than silently accepting partially-implemented behavior.
4. Consider whether the PR should be split into a minimal core endpoint plus follow-up semantics if upstream review stalls on scope.
5. Make sure the PR description clearly distinguishes what is implemented, what intentionally degrades, and what is explicitly unsupported.
6. Only push this aggressively once the higher-priority engine/scheduler correctness items are not in draft, unless local product needs force earlier action.

Acceptance criteria:

• /v1/responses behavior is clearly specified and tested
• unsupported semantics fail explicitly
• the PR is either merged as-is or intentionally split into smaller reviewable pieces

[krystophny]

Fix is on fork main. Tracking upstream merge via the corresponding PR.