Improve local Ollama fallback for low-resource machines

[naraypv]

Summary

This PR improves Graphify's local semantic extraction path for machines that cannot comfortably run large local models.

Implemented changes:

• default local Ollama model policy targets small laptop-safe models: qwen2.5-coder:3b first, then gemma3:4b
• routes Ollama extraction through the native /api/chat endpoint so context options such as num_ctx actually apply
• adds robust fallback behavior so a failed local model can retry the next local model before falling back to MiniMax when configured
• keeps MiniMax as an optional last fallback, not the default workhorse, with controls to cap remote spillover
• adds load/daytime controls and defer mode so background Graphify work is less likely to make a daily-driver laptop lag
• hardens malformed extraction cleanup so bad LLM fragments do not crash graph construction
• updates generated skills/docs and tests for the new fallback policy

Why this may help the community

A lot of users do not have enough RAM/VRAM to run 20B or 30B coding models while also using their machine for normal work. This keeps the local path useful on lower-resource laptops by trying small local models first, failing over cleanly, and only using cloud fallback when the user has configured it. It should make large-project Graphify runs less brittle without forcing everyone into large local model downloads.

Verification

Ran locally before opening this PR:

• uv run pytest -q -> 2537 passed, 28 skipped
• python3 -m compileall -q graphify
• python3 -m tools.skillgen --check
• uv lock --check
• live Ollama fallback smoke: missing primary model retried with qwen2.5-coder:3b and produced graph.json
• graphify update . --no-cluster
• machine credential guard passed for staged changes, outgoing commits, and full HEAD

No repo-specific CONTRIBUTING file or pull request template was present, so this targets the upstream default branch v8 and follows the visible project/test/security expectations.