feat: per-persona model support for swarm deliberation

[NOVA-Openclaw]

Summary

Allow each swarm persona (Strategist, Economist, Naturalist, Skeptic) to use a different Ollama model during forecast deliberation, configured via environment variables.

Motivation

Each persona judges predictions through a fundamentally different lens. Different model architectures have different strengths that align naturally with these roles:

• A reasoning model (e.g. deepseek-r1) excels at the Skeptic's job — stress-testing assumptions, checking base rates, and playing devil's advocate via chain-of-thought reasoning
• A larger general model (e.g. qwen2.5:32b) brings deeper analytical capacity to the Strategist and Economist's geopolitical and macro-economic analysis
• A solid mid-range model (e.g. gemma3:12b) handles the Naturalist's disaster/weather/health pattern matching efficiently

Running all four on the same model means you're either over-provisioning for simple personas or under-provisioning for complex ones.

Configuration

New optional environment variables in .env:

SWARM_MODEL_STRATEGIST=qwen2.5:32b
SWARM_MODEL_ECONOMIST=qwen2.5:32b
SWARM_MODEL_NATURALIST=gemma3:12b
SWARM_MODEL_SKEPTIC=deepseek-r1:14b

When unset, each persona falls back to the main LLM_MODEL, preserving existing behavior. Zero breaking changes.

Changes

• engine/config.py — Add swarm_persona_models dict built from SWARM_MODEL_* env vars
• engine/oracle.py — Add optional model parameter to _complete() (2-line change)
• engine/swarm.py — Look up per-persona model and pass it through to _complete()
• .env.example — Document new env vars with suggested model mappings

Design Notes

• Personas already run concurrently via asyncio.gather, so different-sized models just finish at different speeds — no architectural changes needed
• All models share the same Ollama endpoint (base URL, API key, timeout) — only the model name varies
• The dict comprehension in config filters out empty strings, so unset vars produce a clean empty dict and the fallback path is just model or self.model