fix(ai): honor non-OpenAI model in Runner

[mikemikimike]

Summary/Problem

Fixes #482. The OpenAI Agents compatibility Runner passed detected agents through the OpenAI framework path even when CONDUCTOR_AGENT_LLM_MODEL selected a non-OpenAI provider. With no OpenAI key, the Anthropic examples failed while returning no useful result.

Changes

• Preserve the existing framework bridge for OpenAI-targeted agents.
• Convert OpenAI Agents objects targeting another provider into native Conductor Agents so the configured provider routing is used.
• Preserve OpenAI function tools during conversion and apply the Runner max_turns value.
• Add regression coverage for non-OpenAI dispatch, OpenAI compatibility, and function-tool retention.

Tests

• pytest -q tests/unit/ai/test_openai_compat.py — 3 passed.
• Python 3.12 Linux Docker CI-equivalent: tests/unit — 2608 passed, 1 skipped; tests/backwardcompatibility — 1015 passed; tests/serdesertest — 62 passed.
• python -m ruff format --check src/conductor/ai/agents/openai_compat.py tests/unit/ai/test_openai_compat.py — passed.
• python -m ruff check tests/unit/ai/test_openai_compat.py — passed.
• python -m mypy --no-incremental --explicit-package-bases --follow-imports=skip src/conductor/ai/agents/openai_compat.py — passed with MYPYPATH=src.
• python -m build --wheel --sdist — passed.
• git diff --check — passed.

Compatibility/Known limitations

OpenAI-targeted agents retain the existing framework bridge, while non-OpenAI targets use native Conductor model routing. The full test matrix was validated in a Python 3.12 Linux Docker equivalent; live server integration was not run because it requires CONDUCTOR_SERVER_URL and provider credentials. Remote CI may require upstream approval for fork-originated workflows.

Issue

Closes #482