flymsg

source

Exact Quest and desktop performance work

The accepted exact-fidelity decision controls this work. Neither bounded geometry nor another representation is a substitute for the original selected source. This is an execution ledger, not a claim of performance improvement or hardware qualification.

Completed implementation

Remaining execution gates

  1. Freeze a clean source and verify exact hardware identity, Browser native focus, complete source/context/GPU readiness and unchanged quality. ADB connectivity alone is not readiness. Record transport failures as exclusions.
  2. Profile original versus exact paged skeletons and surfaces separately. Use the existing application CPU stages and bounded passive process/GPU sampling; unavailable counters remain unavailable. Keep readbacks and diagnostic shader substitutions outside cadence windows.
  3. Choose an intervention only after a measured stage identifies its cost. Preserve original Float32 data, primitive order, IDs, colours, transparency and selection. Re-run independent exact geometry/colour/depth/ID checks before timing. If no justified intervention exists, retain the baseline and report that limitation.
  4. Complete 4/16/64 MiB, four-view, three-repeat ABBA admission under the unchanged Wi-Fi protocol. The mandatory negative veto may stop a rejected candidate; it cannot admit a winner with missing views.
  5. Qualify an admitted exact candidate on physical XR: three 120-second sessions per scene, 20 minutes on the heavier scene, ten entry/exit cycles, stereo/head motion/clipping and immersive recovery. Keep the actual-refresh p95/p99 and missed-slot criteria. Human visual observations cannot be inferred from ADB.

Reproduction

Run from clean committed source. All output directories must be new. Exact skeleton comparison uses:

ANDROID_ADB_SERVER_PORT=5038 bg uv run python scripts/quest_compare.py \
  --serial 192.168.1.146:5555 --out /new/exact-skeleton-evidence \
  --scene looming --representation skeleton --keep-awake \
  --unattended --stop-on-regression

Run --representation surface into a different new directory for complete original surfaces. Unattended virtual proximity is temporary and cleanup-verified; its measurements are desktop diagnostics and cannot qualify physical XR. A preparation timeout is not proof that the full configuration is unsupported.

quest_profile.py supplies bounded read-only device samples. Run diagnostic sampling separately from primary cadence comparisons so its overhead cannot asymmetrically favour an arm. Do not alter clocks, Guardian, resolution or MSAA.

Software viewer/context fixtures use check_geometry_pages.py --oracle viewer-skeleton and --oracle viewer-surface. They exercise the real viewer on explicitly declared small fixtures, not the full Quest workload. SwiftShader results establish correctness only. Preserve failures and compare the unchanged baseline before diagnosing a regression.

Delivery boundary

Publish raw evidence, source identities, independent checks, exclusions and cleanup, then gate the exact PR/merge/deploy. Keep #177 open until its actual physical matrix passes. The phone portion of #24, Xbox and scientific evidence are outside this work. A demonstrated Quest hardware limit leaves the affected fluidity requirement unqualified; it never authorizes hidden quality reductions.