Add Ollama provider for self-hosted LLM inference

[The-R4V3N]

Summary

• Adds LLM_PROVIDER=ollama for fully local, zero-cost LLM inference via Ollama
• Uses Ollama's OpenAI-compatible /v1/chat/completions endpoint — no SDK, just native fetch
• No API key required — fully offline after model download
• Configurable base URL via OLLAMA_BASE_URL (defaults to http://localhost:11434)
• Default model: llama3.1:8b (user-configurable via LLM_MODEL)
• 120s default timeout (longer than cloud providers for local hardware)

Changes

• lib/llm/ollama.mjs — new provider following existing pattern
• lib/llm/index.mjs — factory case + export
• crucix.config.mjs — OLLAMA_BASE_URL env var passthrough
• .env.example — updated docs with ollama option
• test/llm-ollama.test.mjs — 11 unit tests (defaults, request format, response parsing, errors, factory)
• test/llm-ollama-integration.test.mjs — integration test (auto-skips if Ollama unavailable)

Closes #30

Test plan

• Unit tests pass: node --test test/llm-ollama.test.mjs (11/11)
• Integration test skips gracefully when Ollama is not running
• With Ollama running: OLLAMA_MODEL=llama3.1:8b node --test test/llm-ollama-integration.test.mjs
• Full sweep with LLM_PROVIDER=ollama generates trade ideas

[calesthio]

I haven't been able to get to this yet because work has been busy, but I definitely plan to review it over the weekend.

[calesthio]

Approved. The Ollama provider wiring looks sound and the main regression checks passed on an isolated review worktree. One non-blocking nit: est/llm-ollama-integration.test.mjs currently only checks whether the Ollama server is reachable, so it hard-fails if the default model isn't installed locally. It would be better to skip unless the requested model is actually present in /api/tags, since the PR description says the integration test should auto-skip when the local setup isn't ready.

[The-R4V3N]

Thanks for the review! Good catch on the integration test — I've pushed a fix (ca8f76c) that now checks /api/tags for the requested model before running, not just server reachability.
If the model isn't installed it'll skip with a descriptive message listing available models instead of hard-failing.

[calesthio]

Lgtm