Optimize global digest lookups

[moacirmoda]

Summary

• replace three global window-ranked revision scans in digest reads with correlated latest-row lookups
• add a reversible index for global catch-up filtering by Discord creation time
• preserve deterministic createdAt desc, id desc latest-event semantics with regression coverage for same-instant ties

Why

The global catch-up query ranked every message, channel, and member revision before applying its guild and time filters. As revision history grows, that makes a small incremental digest pay the cost of sorting unrelated history.

The correlated lookups constrain each latest-row search to its parent entity. The new timestamp index lets selective global catch-ups begin with the messages in range.

Query-plan evidence

Against the exact compiled global digest query on a disposable store with 100,000 synthetic messages, SQLite selected:

• messages_discord_created_at_index (discord_created_at>?)
• message_revisions_message_id_created_at_index (message_id=?)
• channel_detail_revisions_channel_id_created_at_index (channel_id=?)
• member_detail_revisions_member_id_created_at_index (member_id=?)

Removing the new index changed the outer access path from one global time-range lookup to repeated channel/time lookups through the pre-existing compound index.

Verification

• pnpm run lint
• pnpm run tsc
• pnpm run test:unit twice: 589/589 passed each time
• pnpm run test:e2e: 58/58 specs passed; 20/20 tools reached through real MCP traffic
• pnpm run test:seed-coverage: 17/20 tools demonstrated; the existing 3 live-Discord tools remained explicitly unseedable
• migration verified with migrate -> rollback -> migrate on a disposable store
• focused real MCP stdio session verified messages_catch_up after the final commit