test(translation): table-driven tests for Anthropic <-> OpenAI <-> Responses translation

[Destynova2]

Summary

• Add 29 dedicated tests in tests/enterprise/translation_test.rs covering the Anthropic Messages <-> OpenAI Chat Completions <-> Responses API (Codex CLI) translation paths. Pins the wire format of the three transformers that account for the bulk of inbound and outbound format work.
• Expose a pub mod test_api shim in src/providers/openai/mod.rs that re-emits the otherwise pub(crate) transform_request, transform_response, and transform_to_responses_request as JSON-returning helpers, so the tests/ integration suite can reach them without enabling extra cargo features.
• Audit context: 18 touches in src/server/openai_compat/, src/server/responses_compat/, and src/providers/openai/transform.rs over 3 months with no dedicated tests. This file is the regression ratchet that catches structural drift before release. Streaming is intentionally out of scope here (parallel PR).

Test matrix

Anthropic Messages -> OpenAI Chat Completions (request direction):

• anthropic_to_openai_simple_user_message
• anthropic_to_openai_with_system_prompt_hoisted_to_first_message
• anthropic_to_openai_strips_duplicate_system_role_from_messages
• anthropic_to_openai_with_assistant_message_after_user
• anthropic_to_openai_tool_use_block_to_tool_calls
• anthropic_to_openai_tool_result_block_to_tool_role_message
• anthropic_to_openai_image_input_url
• anthropic_to_openai_image_input_base64
• anthropic_to_openai_thinking_block_dropped_when_target_is_chat_completions
• anthropic_to_openai_cache_control_dropped_when_target_lacks_cache
• anthropic_to_openai_with_max_tokens_and_stop_sequences
• anthropic_to_openai_temperature_and_top_p_passthrough

OpenAI Chat Completions -> Anthropic (response direction):

• openai_response_finish_reason_stop_to_anthropic_end_turn
• openai_response_finish_reason_length_to_anthropic_max_tokens
• openai_response_finish_reason_tool_calls_to_anthropic_tool_use_block
• openai_response_usage_to_anthropic_usage_input_output
• openai_response_with_no_content_to_anthropic_with_empty_text_block

Inbound OpenAI request -> Canonical (server compat layer):

• openai_request_to_canonical_simple_user_message
• openai_request_to_canonical_with_image_url_part
• openai_request_to_canonical_tool_call_followed_by_tool_result

Canonical response -> OpenAI response:

• canonical_response_to_openai_finish_reason_stop

Anthropic Messages -> Responses API (Codex CLI):

• anthropic_to_responses_input_array_format
• anthropic_to_responses_with_tools_array
• anthropic_to_responses_streaming_event_types_match_spec (static type-tag pin only; streaming codepath untouched)

Edge cases:

• transform_returns_error_on_unserializable_tool_input
• transform_handles_empty_messages_array_gracefully
• transform_handles_unicode_in_message_content
• transform_handles_very_long_messages_no_panic (10MB string)
• transform_preserves_request_id_in_response_metadata

Test plan

• cargo check --tests passes
• cargo fmt --all -- --check passes
• cargo clippy --tests -- -D warnings passes
• cargo nextest run --test lib enterprise::translation_test passes (29/29)
• Full lib suite still green (246/246)
• All pre-push hooks pass locally (gitleaks, fmt, clippy, deny, machete, doc-coverage, audit, sweep)

🤖 Generated with Claude Code