[amxson]

Description

Fixes #408.

The OpenAI Responses converter used the response item's id as the native tool-call correlation ID, losing the distinct call_id required to pair a function call with its result. Non-function output items such as reasoning state were also discarded before the immediate follow-up request.

This change:

• preserves the original Responses call_id through native tool execution and the corresponding tool result;
• carries provider-native response output as optional opaque state for the immediate follow-up;
• replays the original Responses output in order without duplicating function calls;
• leaves other providers unchanged when no native replay state is present.

This does not change multi-turn history retention or provider error/lifecycle behavior from #409.

How Has This Been Tested?

• Added focused OpenAI Responses regression coverage for call correlation, reasoning replay, parallel calls, and generic-provider fallback.
• New regression tests: 4 passed.
• Focused provider/budget/timeout tests: 71 passed.
• Provider model tests: 27 passed.
• Full Python suite: 246 passed, 3 skipped.
• MeTTa suites: 6 passed.
• Mandatory suite: 243 passed.
• Native-loop probe verified one function call, preserved reasoning state, and correct call_id round-trip.
• Original v0.1.21 OpenAI Responses converter loses call_id and opaque reasoning replay #408 fixture now returns:
  {"call_ids": ["call_actual_73", "call_actual_73"], "reasoning_replayed": true}
• git diff --check: passed.

Checklist

• PR contains autogenerated code
• Self-review completed
• Test scenarios above are passed with the version of the code from PR