Problem

memory.extended.llm was all-or-nothing: without both base_url and model, ResolveLLM fell back to the main client entirely. Operators could not override a single knob — e.g. disabling thinking for memory calls — without duplicating the parent's base_url/api_key/model.

Change

ResolveLLM now inherits empty fields (base_url, api_key, model, thinking, max_tokens, temperature) from the main *llm.Client before building the dedicated client. Partial overrides now work:

"llm": { "thinking": "disabled" }   // same model/endpoint, no reasoning for memory calls
"llm": { "model": "qwen2.5:7b" }    // cheaper model on the same backend

Fallback semantics are preserved: if base_url/model are still empty after inheritance (e.g. the main client is not an *llm.Client), it warns and reuses the main client as before. timeout_seconds is not inherited (the main client's timeout is unexported) and still defaults to 30s.

Docker setup

Both docker configs (config.godmode.json, config.restricted.json) now set memory.extended.llm.thinking: "disabled" via the new partial override, so memory extraction/rerank/predictor calls skip reasoning tokens while reusing the parent backend.

Testing

• TestResolveLLMInheritsMainSettings — thinking-only override inherits base_url/api_key/model/max_tokens/temperature
• TestResolveLLMPartialOverrideKeepsInheritedRest — model-only override keeps the parent backend
• Existing fallback tests unchanged and passing; full internal/memory/... suite green with -race

Docs updated: docs/CONFIG.md (memory.extended.llm row) and docs/EXTENDED_MEMORY.md (Dedicated Memory LLM section).