fix(translation): fail closed on malformed Responses input

[leseb]

Summary

The Responses-to-Chat Completions translation silently dropped or synthesized malformed input (scalar input became an empty message list; non-object items were skipped; missing message role/content were defaulted; function calls/outputs missing call_id, name, arguments, or output were dropped), so the backend received a different, valid-looking Chat Completions request instead of the client's invalid Responses request being rejected. This change validates translation-required structure before canonical state erases the original scalar input and routes unsupported input types and missing required item fields through the existing TranslationError → HTTP 400 path, leaving valid requests byte-for-byte unchanged.

Related issue

Closes #552

Validation

• Unit tests — cargo test -p praxis-ai-apis (2,589 passed, +2 doctests)
• Integration or functional tests — cargo test -p praxis-tests-integration responses_to_chat_completions (7 passed, incl. new responses_to_chat_completions_rejects_malformed_input_before_upstream, which asserts the backend receives no request)
• make lint
• Additionally: make build, git diff --check

Checklist

• I reviewed every changed line and can explain the change.
• New capabilities include an example config and functional example test. — N/A: hardening of an existing capability; reuses responses-to-chat-completions.yaml with a new functional test.
• User-facing behavior and generated documentation are updated. — Malformed input now returns HTTP 400 instead of being silently altered; no generated docs affected (make lint doc-sync checks pass).
• Performance-sensitive changes include appropriate benchmark or load-test evidence. — N/A: validation on the existing request-parse path; no new buffering or clones.
• Commits are signed and include a Signed-off-by trailer.

Breaking changes

Requests with malformed Responses input that were previously silently repaired now return HTTP 400 with an invalid_request_error. Well-formed requests are unaffected. This aligns the proxy with the Open Responses schema; any client relying on silent repair must send schema-valid input.