Evaluate focused agent context against the v0.4 baseline #53

Goal

Test whether a stricter agent-read context strategy fixes the three small-page over-selection failures found by the offline baseline while preserving target recall and large-document savings.

Scope

• Add an experiment-only context strategy interface.
• Keep the current full-outline and targeted-planner measurements unchanged.
• Add a focused candidate that uses explicit block/component targets, ambiguity, missing refs, and bounded neighbours without automatically expanding the entire section.
• Report candidates side by side against the same frozen cases and gates.

Guardrails

No product-context change and no public MCP API change in this issue. Promote the strategy only after the experiment shows equal target recall, zero forbidden selections, and useful context reduction.