What

This PR teaches lcm_expand_query to answer one focused question across multiple conversations when allConversations: true is set. It groups globally matched summaries by conversation, expands a bounded set of ranked conversation buckets sequentially, and returns one merged answer with per-conversation diagnostics. It also tightens prompt guidance so agents prefer local scope first and only escalate to global search/synthesis when the current summaries appear insufficient.

Why

Lossless recall is most useful when work spans multiple sessions, but the old flow could only expand one conversation at a time even after global discovery. That forced agents to manually run repeated expand calls and stitch answers together. This change makes cross-session synthesis practical while keeping delegation bounded and explicit.

Changes

• Add ranked multi-conversation bucket planning
• Expand top conversation buckets sequentially
• Merge delegated answers into one response
• Report plural conversation scope diagnostics
• Cover partial, skipped, and budgeted runs
• Clarify local-vs-global recall prompt guidance
• Add a minor changeset for release notes

Testing

• pnpm vitest run test/lcm-expand-query-tool.test.ts
• pnpm vitest run test/plugin-prompt-hook.test.ts
• Expected: both test commands pass