[codex] Respect explicit Ollama think setting

[TheEditor]

Summary

Respects explicit think settings passed to OllamaClient.chat() while preserving the current default behavior for reasoning models.

When callers omit think, reasoning models such as qwen3 and deepseek-r1 still default to think=true. When callers pass think=false or think=true, that explicit value is now forwarded unchanged to Ollama.

Why

Some bounded calls deliberately disable reasoning, especially smoke tests, exact-answer probes, routing checks, and short command-style requests. The previous behavior overrode think=false for reasoning models, which could consume a small num_predict budget on hidden reasoning and return no visible streamed content.

This keeps the reasoning-model default intact while allowing callers to opt out when the call shape requires it.

Changes

• Only auto-enable think=true for reasoning models when think is omitted.
• Preserve explicit think=false for reasoning models.
• Preserve explicit think=true for non-reasoning models.
• Add tests covering the default and explicit override behavior.

Validation

• PYTHONPATH=SourceCode python3 -m unittest tests.test_ollama_wait_for_available tests.test_ollama_stream_watchdog tests.test_inference_router_compat
• Live local deployment smoke: qwen3:14b through InferenceRouter with think=false, num_predict=48, and num_ctx=8192 returned visible content instead of empty streamed content.
• git diff --check