Quest Wi-Fi observer trials

This protocol implements the accepted observer-rendering decision. It does not qualify a renderer or change the selected anatomy. Run from clean committed source with the Quest connected through its explicit TCP ADB serial. Keep USB disconnected from the test transport; there is no USB recovery fallback.

ANDROID_ADB_SERVER_PORT=5038 uv run python scripts/quest_compare.py \
  --serial 192.168.1.146:5555 --out runs/quest-wifi-comparison \
  --scene looming --representation skeleton --keep-awake

The wearer keeps Quest Browser foreground and does not enter VR during this desktop comparison. Ordinary keep-awake is temporary; presence sensors, display resolution, GPU clocks and browser rendering flags are unchanged. Each owned tab, forward, reverse, server and keep-awake setting receives an independent cleanup attempt. cleanup.json records failures; device.json retains the original setting for Wi-Fi recovery. A failed restore never excuses skipping subsequent cleanup steps.

Preparation permits up to 600 seconds for complete selected anatomy and GPU context. A timeout is an excluded preparation, not evidence that the scene is unsupported. Source attestation uses Debugger before warmup and after measurement only. During timing, callbacks and submissions come from the existing application animation loop. All intervals survive, including long stalls and windows longer than 600 frames. Readbacks, GPU inventory and correctness checks use separate pages.

The independent original renderer supplies colour, depth and neuron IDs at 0, 20 and 350 ms for two eye offsets. Primitive spans must cover every original selected primitive exactly once; full-quality desktop captures must also agree. A candidate that fails this gate is not timed. Fresh documents prevent readback work from contaminating the cadence arms.

For each of 4, 16 and 64 MiB pages, run A/B/B/A, where A is original and B paged, for 30 seconds per arm after three seconds of warmup. Repeat three times in each of dense-all, anatomical-motion, close-motion and close-static. Dense-all disables culling. Motion uses the application loop timestamp with yaw amplitude 0.15 rad and period 10 seconds; close views scale the frozen camera position by 0.35. Replay stays at 20 ms. Geometry, viewport, source and image quality must agree. A failed measurement invalidates its entire block, with at most three invalid blocks per page configuration. Failed preparation stops that configuration.

Admission requires at least 10% median p95 gain in one motion view with a favorable sign in every repeat, and no view with over 10% median regression. Raw intervals are independently summarized in Python. Within one percentage point, measured GPU memory breaks the tie. comparison.json retains every result and admission reason. No admitted candidate means no XR performance claim. It does not establish that all Quest configurations are unsupported.

Repeat surface admission separately, without replacing missing surfaces by skeletons. Validate any winner against the original anatomy scene if that complete scene is admissible. Neither a partial scene nor another representation substitutes for these gates.

Only an admitted candidate proceeds to the assisted physical XR window: three 120-second sessions per admitted scene, then 20 minutes on the heavier scene. Check stereo, scale, clipping and head movement; ten enter/exit/menu cycles; context recovery and suspend/resume. Use scripts/quest_xr.py --wifi-only --check with the winning page size and retain session reports. Source attestation finishes before READY; subsequent device sampling is passive. The document identity in each XR report must match the attested document. Existing qualification thresholds remain p95 <= 1.1 actual budgets, p99 <= 2 budgets and missed-slot estimate <= 1%; unknown or mixed refresh rates are indeterminate. Record human observations explicitly; automated timing cannot certify stereo comfort or menu responsiveness.

These commands are collection infrastructure. Hardware qualification remains open until complete, valid device evidence and the physical checks exist. Scientific validation, Xbox hard-stop/GPU-fault recovery and surface admission remain separate.

Optional mandatory-veto stopping rule

After discussing the measured collection cost with the owner on 2026-10-16, use --stop-on-regression to stop a candidate after three valid repeats in any one view establish a median ratio above 1.10. This is the existing mandatory rejection criterion. It can reject a candidate with explicitly unmeasured views; it cannot admit one. Positive admission still requires every repeat of every view. The first 4 MiB campaign used this amendment after its three dense blocks completed; its interruption and independent veto record remain alongside all raw samples. Subsequent configurations record the option in their protocol before collection.

Unattended diagnostic amendment (2026-10-16)

The operator requested collection without wearing the headset. quest_compare.py --unattended uses the temporary virtual proximity interface documented by Meta: https://developers.meta.com/vr/documentation/unity/meta-xr-operator/quest/ It refuses existing proximity overrides or wake leases, registers restoration before mutation, and verifies return to physical sensor control during cleanup. All transport remains explicit Wi-Fi. XR query parameters are omitted and XR receipts are ineligible. These campaigns are desktop diagnostics only; do not combine their measurements with physical-presence campaigns or claim normal XR, comfort, head-motion, or stereo qualification. Rendering quality stays unchanged. The initial device roundtrip held Awake for 25 seconds and verified restoration.

Static readback preparation

quest_bounded.py explicitly requests readback_only readiness. A complete, visible, awake static page can legitimately provide too few animation callbacks for the cadence preflight. This separate path still checks URL, document, loaded sources, full anatomy, context and GPU state; it requires completed nonuniform readback images and actual line counts in the experiment. It cannot be combined with required rendering progress or an immersive session, and never sets an XR-eligible document receipt. Ordinary cadence/XR collectors retain their callback and submitted-frame checks unchanged. Excluded image attempts are retained alongside the successful diagnostic, not rescored as timing data.