[deepanwadhwa]

Summary

• send two image attachments through one native Molmo inference with labelled visual inputs, a shared crop budget, and sequential vision encoding
• reject unsupported third-image requests explicitly instead of splitting or silently degrading them
• hand grounded Molmo evidence back to the selected text model for a bounded final synthesis while preserving direct Molmo behavior
• strengthen deep-file retrieval prompting and coverage
• disclose response-limit truncation beside partial UI answers
• retry the transient macOS launchd bootout/bootstrap race with deterministic lifecycle coverage
• update Molmo usage, installation, model, and qualification documentation

Local validation

• make
• make omp
• make test
• make test-molmo2
• node tests/test_web_activity_ui.mjs
• git diff --cached --check

This PR contains the single follow-up commit 35aaee1; the previous LAN work is already present on main.