Support local query embeddings with visible fallback

[plasmaStray]

What changed

• route query-time dense embeddings through configurable EMBEDDING_API_BASE, EMBEDDING_MODEL, and optional EMBEDDING_API_KEY
• keep query embeddings synchronous so Icarus finishes retrieval before Hermes calls the main LLM
• add configurable timeout and retry handling for local model cold starts
• surface degraded dense retrieval to the user through trusted pre-LLM warning context while preserving lexical/session fail-open behavior
• retain backward-compatible OpenRouter defaults and the list-only helper API

Why

The worker already supports OpenAI-compatible local embedding endpoints, but scripts/context_enhancer.py hardcodes OpenRouter for query embeddings. That makes local indexing and query-time semantic retrieval asymmetric and silently disables dense Wiki recall for installations without an OpenRouter key.

User impact

Operators can use the same local embedding model for indexing and queries. A cold local backend is awaited before prompt construction. If dense retrieval times out or fails, the assistant is instructed to disclose that Wiki/context recall may be incomplete instead of silently answering as if semantic search succeeded.

Checks

• python3 -m py_compile scripts/context_enhancer.py icarus/hooks.py _test_local_embeddings.py
• python3 _test_local_embeddings.py
• python3 _test_sanitize.py
• python3 _test_collapse.py
• bash -n setup.sh
• git diff --check