[deepanwadhwa]

Summary

• land the durable deep-file chat plan and native VisionPsy/Molmo2 multimodal runtime already developed on this branch
• add password-protected LAN access through a simple local URL while keeping all model inference and chat processing on the host Mac
• keep raw model backends loopback-only, add per-launch browser sessions, and preserve same-origin protection
• document concurrency, global model selection, security boundaries, startup, troubleshooting, and shutdown behavior
• include runtime safety, app supervision, model handoff, UTF-8, and Maple token-limit fixes from the existing branch history

LAN usage

Run samosa app --lan, open the printed local-network URL on another device, and sign in with the configured password. The temporary default is password1234 and can be changed with SAMOSA_LAN_PASSWORD.

Validation

• make
• make omp
• full make test, including the real non-loopback LAN test
• exact-index snapshot: compiled gateway and fake backend, shell syntax, wrapper fixtures, browser UI fixtures, and LAN integration test
• git diff --cached --check

Review note

The LAN implementation is isolated in commit b036a9c. Uncommitted Molmo multi-image and launchd retry work remains local and is not part of this PR.