Build an offline routing and agent-context experiment harness #51

Goal

Measure whether a lightweight routing layer and narrower agent context improve Lucid before either is added to the runtime.

First executable slice

• Freeze routing and context-selection cases by intent family.
• Add a shared candidate contract and the current Lucid baseline adapter.
• Compare full-outline context with the existing targeted context planner.
• Emit deterministic JSON and Markdown reports with safety and value metrics.
• Keep candidate models outside the product runtime until they pass the experiment gates.

Guardrails

No product routing change, no public MCP API change, and no model dependency in this slice. The existing parser/compiler, revision, destructive-change, and undo safeguards remain authoritative.