fix(reflect): emit canonical tool result messages

[Jordan-Jarvis]

Summary

• remove the non-standard name field from Reflect role=tool messages
• preserve the canonical OpenAI tool-result shape at the shared Reflect producer boundary
• derive Gemini FunctionResponse.name from the immediately preceding assistant tool calls by tool_call_id
• reject unknown, duplicate-within-turn, and missing tool results while allowing providers that reuse call IDs across turns

Root cause

Reflect currently adds name to tool-result messages for Gemini. Strict OpenAI-compatible gateways reject that extra field because tool-result messages permit content and tool_call_id; provider-specific requirements should be handled in the provider adapter instead of changing the canonical message shape.

Verification

• targeted Ruff on all touched files
• pytest -q tests/test_reflect_agent.py tests/test_gemini_safety_settings.py -k 'not RealLLM and not ContextOverflowIntegration' (69 passed)
• repository pre-commit hooks and lint passed
• git diff --check

No second Reflect path or compatibility fallback is introduced.

[Jordan-Jarvis]

Superseded by #2766, which carries the same canonical producer/adapter fix with the stricter unknown-ID and duplicate-within-turn validation coverage. Keeping one upstream contribution avoids split implementations and duplicate review.