Validate chat completion requests before upstream fetch

[xodapi]

Why

The proxy currently accepts arbitrary JSON for /v1/chat/completions after parse. A small validation layer would catch malformed requests early and reduce confusing upstream failures.

Scope

Validate a conservative OpenAI-compatible subset:

• messages must be an array when present;
• model must be a string if present;
• numeric limits such as max_tokens and temperature must be bounded when present;
• reject obviously unsafe or unsupported shapes with a 400 response.

Acceptance criteria

• Invalid JSON still returns 400.
• Malformed chat body returns 400 before upstream fetch.
• Valid OpenAI-compatible requests still pass through.
• Tests ensure no prompt/response data is logged or persisted during validation errors.