Add Responses translation semantics

[franciscojavierarceo]

Part of #213.

Summary

• extend the translator with schema-complete Responses resource and stream-event conversion
• preserve function-call history, structured output formats, allowed tool choices, strict function schemas, logprobs settings, and content_filter incomplete status
• add curated Chat Completions recording fixtures; the fixture provenance comes from OGX compatibility testing

Stack

• depends on Share Responses body size limits #701
• depends on Add Responses request translation core #702
• base of Add Responses to Chat Completions filter #696
• followed by Add Responses SSE framing coverage #677

Validation

• cargo test -p praxis-proxy-filter --lib --features ai-inference translation
• cargo test -p praxis-proxy-filter --lib --features ai-inference openai::responses
• make lint

[praxis-bot]

Review: Add Responses translation semantics

Extends the Chat Completions translator with response/stream conversion and schema-complete resource building. The request-side additions (logprobs, text format, stream options, function-call history, strict schemas) are well structured and tested. The response and stream translators are substantial and mostly correct.

Severity	Count
Medium	3

Non-inline findings

None -- all findings are inline.

[praxis-bot]

[Medium] append_function_call creates a separate assistant message per function call item. When the Responses input contains consecutive function_call items (parallel tool calls), the Chat Completions API expects them batched into a single assistant message's tool_calls array. Separate messages will cause provider errors because the intervening tool response messages won't have a matching preceding tool_calls entry.

Merge consecutive function_call items into one assistant message, or accumulate all tool calls and emit a single message at the end of the function-call sequence.

[praxis-bot]

[Medium] The base version fell back to part.get("file_id") for input images, which matches the Responses API input_image schema. This PR changes the fallback to part.get("url"), which is not a documented input_image field. If a caller sends {"type": "input_image", "file_id": "file-abc"}, it will now fall through to the else branch and clone the raw part instead of extracting the file reference.

Restore file_id as an additional fallback: .or_else(|| part.get("file_id")).

[praxis-bot]

[Medium] process_stream_choice only extracts delta.content (text). Tool-call streaming chunks carry deltas in delta.tool_calls[].function.arguments and delta.tool_calls[].function.name, which are silently dropped here. A streaming request that triggers function calls will produce a terminal response.completed event with an empty output array and no function call items.

If tool-call streaming is out of scope for this PR, add a // TODO or tracking issue and document the limitation. Otherwise, accumulate delta.tool_calls fragments and emit function_call output items in finish_stream.