ollamaUrl was built as "http://" + OLLAMA_HOST, so any value carrying a scheme
- the form Ollama's own tooling accepts - was mangled:
OLLAMA_HOST=https://ollama.example.com
-> http://https://ollama.example.com
The LLM attack then could not connect, and reaching a remote Ollama over TLS
was impossible. llm.py derives its client base URL as f"{ollamaUrl}/v1", so the
malformation propagated into the OpenAI-compatible client.
Extract _normalize_ollama_url() so the logic is unit-testable - ollamaUrl is
assigned at module import, which is otherwise only reachable via reload - and
prepend the scheme only when absent. Also strip trailing slashes, since callers
append paths. The bare host:port default is unchanged.
Fixes#119
Co-Authored-By: Claude <noreply@anthropic.com>
