feat(llm): enable thinking for Ollama via default additional params

[standardtoaster]

Summary

Ollama's native /api/chat requires an explicit "think": true to enable extended reasoning. Without it, thinking-capable models (Qwen3, DeepSeek-R1, Gemma 4) run without their reasoning step.

This PR always sends think: true for Ollama. Non-thinking models ignore the parameter harmlessly -- Ollama treats it as a no-op. This avoids maintaining a hardcoded list of thinking-capable model names that would rot as new models ship.

The implementation adds a generic with_additional_params() builder on RigAdapter for injecting provider-level defaults into every request. The Ollama factory uses it to set think: true. Other providers are unaffected.

Design decisions

Why not a per-model config flag? IronClaw doesn't have per-model config today. Adding one for a single boolean isn't worth the complexity. If Ollama adds models where think: true causes problems, we can add an opt-out then.

Why not a hardcoded model pattern list? Model lists rot. New thinking models ship regularly (Gemma 4 just launched). Ollama handles the parameter gracefully for non-thinking models, so there's no safety concern.

Why with_additional_params instead of a think-specific flag? The generic approach is reusable -- any provider can inject top-level fields without new RigAdapter API surface per feature.

Test plan

4 unit tests covering merge_additional_params:

• Merge into empty params
• Merge preserves existing keys (e.g., cache_control)
• Existing keys win over defaults (no silent overwrite)
• No-op when defaults are None

Manual wire-level validation -- started IronClaw with Ollama backend pointed at a mock HTTP server, captured the raw /api/chat request body:

BEFORE (base): "think": false   (Ollama default)
AFTER  (fix):  "think": true    (injected by RigAdapter)

[gemini-code-assist]

Code Review

This pull request introduces the ability to provide default additional parameters to the RigAdapter, which are merged into every completion request. Specifically, it enables the think: true parameter for the Ollama provider to support extended reasoning in thinking models. The implementation includes a new with_additional_params method, a merging utility that preserves existing request parameters, and corresponding unit tests. I have no feedback to provide.