fix: fall back to Ollama native /api/chat for thinking-mode models (fixes #26)

[nandanadileep]

Problem

Thinking-mode models (e.g. gemma4:26b) generate internal <|think|> reasoning tokens that exhaust max_tokens before producing visible content. Ollama's OpenAI-compatible /v1/chat/completions endpoint strips those tokens and returns empty content, causing 500 errors when starting simulations.

Fix

In LLMClient.chat(), after calling the OpenAI-compat endpoint, check if content is empty. If it is and we're talking to an Ollama server, retry via the native /api/chat endpoint, which surfaces the visible response correctly.

Changes in backend/app/utils/llm_client.py:

• Added _ollama_native_base() — strips /v1 suffix to get the Ollama host URL
• Added _chat_via_ollama_native() — POSTs to /api/chat with stream=false, carries over temperature and num_ctx
• In chat(): triggers the fallback only when content is falsy and _is_ollama() is true — fully backwards-compatible, zero impact on non-Ollama or non-thinking-mode models
• Fixed a latent NoneType crash: re.sub(…, content or '') guards against None content even without the fallback

Test plan

• Run a simulation with a standard model (e.g. qwen2.5:32b) — behaviour unchanged
• Run a simulation with gemma4:26b — should now return visible response instead of 500
• Verify _ollama_native_base() strips /v1 from http://localhost:11434/v1 correctly
• Non-Ollama endpoints (OpenAI, etc.) are unaffected — fallback never fires

Fixes #26