fix: align sandbox smoke with Ollama

[cgfixit]

What

• point the CyClaw sandbox mock at the current Ollama endpoint (127.0.0.1:11434/v1)
• use the configured Ollama model id (qwen2.5:7b)
• temporarily rewrite models.local_llm as well as Grok/Claude during the smoke run
• update the health contract and sandbox guidance from lm_studio to ollama
• add regression coverage for the local provider rewrite

Why

main migrated the runtime to Ollama, but the sandbox harness still mocked LM Studio on port 1234 and only rewrote external providers. That meant the smoke test could exercise a real local Ollama instance—or fail for environmental reasons—instead of the deterministic mock it claimed to verify.

Verification

• python -m pytest tests/test_cyclaw_sandbox_skill.py -q --tb=short — 4 passed
• ruff check --select E,F,I,B,C4,UP,S ... — clean
• python -m py_compile ... — clean
• invariant guard — 27 passed, 0 failed
• dependency guard — 0 failures, 0 warnings
• doc-sync guard — 0 drift
• git diff --check — clean

Risk

Low. The changes are isolated to the repo-local sandbox skill, its mock, documentation references, and focused tests. config.yaml remains restored after each run.