Start with conversational missions and add held-object throwing

[axelpey]

First arrival offers three prepared missions: put away a LEGO brick, find the Backrooms exit, or cross the street. IntroAgent guides ordinary conversation using map-matched spatial memories. Only the physical mission result or explicit Skip unlocks the interface; reopening resumes the same attempt. Completed/skipped first runs persist on the stable broker origin. Help is a separate, responsive interface tour, and completing/skipping a first mission suppresses the redundant Challenges introduction.

Adds ThrowObject for a small object already held, with synchronized gripper release and grasp checks before the committed swing. Preserves wrist roll, guides the agent to approach within short-toss range, and tells it to finish explicitly requested action sequences. Pickup retention is checked after carry. Misses remain recoverable through chat.

Stacked on draft #721 (based on #757). Includes OS environment bridge #776; paired broker controls/completion persistence are in innate-cloud#110. The local combined preview also includes CTA change innate-cloud#107. Durable LEGO pickup is a separate dependency: local integration 2af6ae82e borrowed the validated Astra candidate based on 0a19eff and OpenAI transport from #772; those files are NOT part of this PR. Successful pickup→approach→throw evidence is preserved at local 7ed351fc3. The pickup speed task is still revising its candidate, so one final compatibility run is pending before the combined work is ready.

Validation: all nine webapp test files pass; focused challenge/memory/intersection physics tests pass; three ROS-backed throw tests pass; viewer environment tests, build and typecheck pass. Full webapp typecheck still has existing errors outside these changes. Live desktop/mobile chooser, conversational retry, reopen during pickup, Skip, keyboard handling and passive Help verified. All three missions passed real physical judges. The cleanup retry completed pickup→30 cm approach→throw and settled the brick inside the box. A fresh first-run street pass revealed the UI automatically; completion survived a fresh iframe origin, without a second introduction. Tests also caught a mistaken agent success claim: it did not unlock the interface, and normal chat correction continued the mission. Agent distance/destination judgments can still need user correction. No physical robot testing or deployment.