Refactor browser capture to inference scheduler

[hexadecimal-REBOOT]

Summary

• replace the worker-driven capture pipeline with an in-browser inference scheduler that runs depth every frame and decimates YOLO
• update the UI orchestration to reuse propagated boxes/depth, simplify recording, and drop the unused worker scripts
• align single-frame and recording VDSX exports with the new scheduler data, including updated provenance strings

Testing

• not run (not requested)

---

https://chatgpt.com/codex/tasks/task_e_68e2d4c3bd70832a84d28f90c4efd5de