Agent over-thinks simple Galaxy tool-availability questions until the user has to abort

### Observed
Asking a direct, answerable question like "What tools are available and installed on Galaxy to annotate viral genomes?" causes multiple models to think for a very long time -- long enough that the user gives up and aborts the session. That question maps cleanly onto a single Galaxy MCP lookup (`search_tools_by_keywords` / `search_tools_by_name`), not something that should require extended reasoning.

### Expected
For straightforward tool-availability / catalog questions, the agent should route quickly to the Galaxy tool-search MCP and answer, rather than burning a long thinking budget before (or instead of) calling the obvious tool.

### Hypothesis / levers (both loom-side)
1. The always-on system prompt + full tool surface may be too heavy for the model to dispatch quickly -- the territory of #99 (right-size the system prompt + tool surface for smaller/open models).
2. There may be no steering nudging "for tool-availability questions, just call `search_tools_*`." `extensions/loom/context.ts` already references tool search, so a concise routing hint could short-circuit the over-reasoning.

Recommend linking #99 as the systemic fix and considering a targeted steering nudge for catalog/availability questions as the cheap mitigation.

### Environment
Orbit 0.3.1, macOS arm64, provider google / gemini-3.5-flash. Reported as affecting "multiple models," so likely not provider-specific.

Start by reading extensions/loom/context.ts and the routing around the Galaxy MCP functions search_tools_by_keywords and search_tools_by_name. Compare the issue with #99 and determine whether a targeted steering hint is appropriate; done means straightforward tool-availability questions reach the tool search promptly without extended reasoning, with relevant behavior covered by the project's existing validation.

Written by the indexing model from the issue text.