Sampling tweaks fix Qwen 3.5 overthinking

Qwen 3.5's overthinking issue highlights a growing need for "reasoning constraints" in local LLMs to prevent efficient models from becoming slow liabilities.

• –**Tool-Induced Brevity**: Activating tools swaps the model's dense bullet-point reasoning for a more concise "Claude-style" internal monologue.
• –**Presence Penalty is Key**: Increasing `presence_penalty` to 1.5 helps prune the repetitive tokens that lead to infinite loops in long-context reasoning.
• –**Workflow-Dependent**: The fix is particularly effective in Open-WebUI with native function calling enabled, though it applies to any harness that can handle tool definitions.
• –**Community-Driven Refinement**: This reinforces the trend of local LLM power users having to prompt-tune base models to reach their full potential without switching to larger variants.