[superposition]

Camera readers, actuator acknowledgements and the host cognition boundary were preventing the Qualia integration from using live robot state. This change shares one live V4L2 camera stream, adds measured automatic lighting and a separately authorized camera-aim endpoint, correlates zero-drive completion with the exact verified safety receipt, and publishes the existing L2 belief state in the canonical contract understood by the host.

The acknowledgement repair retains the 500 ms timeout and exact command/request correlation. Nonzero acknowledgements, the deadman, verified stop and wheel authority gates remain unchanged. Camera-aim authority grants no wheel authority. Cognition adaptation preserves latent values, timestamps, sequences and digests; return frames require authentication, expiry and replay checks.

Validation: full eight-feature AArch64 production builds with CUDA; actual camera/light readbacks with operator confirmation of illumination; zero drive HTTP 200 followed by verified stop and lease closure; advancing host L3?L6 sequences and changing activation/weight sketches after boundary deployment. Host upper layers currently use CPU fallback; robot CUDA remains active. Return-channel acceptance is unconfirmed because the protected Windows credential comparison was canceled. No nonzero-wheel, deadman-under-motion, estop/resume or learned-driving demonstration is claimed. Evidence is under docs/evidence; no credentials or camera imagery are committed.

[logohere]

Review disposition: keep draft, do not merge as ready. The rust job is failing cargo fmt --check (run 34733421510); passing feature-matrix jobs do not override that failure. The body also leaves return-channel acceptance unconfirmed. Reconcile remaining changes against current main, rerun current-head checks and resolve the actual acceptance evidence before requesting merge. No nonzero-motion or end-to-end learned-control proof is inferred.