Observer rendering qualification

The owner accepted the rendering ADR on 2026-10-15. Quest 2 remains unqualified. The initial immersive timing result failed; the adaptive-controller fixes did not establish that the original workload fits. The current diagnosis includes a real-device trial with every original skeleton loaded and rejected shader experiments. Performance remains unresolved; this work must not be described as a lag fix.

The 2026-10-16 Wi-Fi performance follow-up rejects the measured 4 MiB dense candidate and records the owner-accepted half-pixel skeleton experiment. That separate fixed-view experiment improves serialized desktop work; it is not production integration or XR qualification.

Implemented contracts

• representation=surface (default) requests original full-resolution surfaces. Admission is all-or-none, using the existing compressed-size estimate followed by actual buffer capacity checks. Estimates are not GPU memory measurements. Missing data, incomplete loading and exhausted capacity are visible states. Unsupported content is never substituted with skeletons.
• representation=skeleton explicitly requests every selected neuron's original skeleton, without geometric simplification. It does not allocate surface geometry buffers or queue surface manifests. Neither strict representation activates adaptive pixel-ratio reductions.
• representation=mixed reproduces the old diagnostic selection policy. Its automatic substitutions are explicitly unqualified; it is not an accepted fidelity mode or a successful fallback.
• pages=1 enables the candidate spatial scheduler. Workers assign original segments to immutable pages. Every endpoint remains unchanged, crossing edges belong to exactly one page, and compaction retains original draw order. Complete strict skeleton scenes use an immutable resident original-order index arena and one compact-view cache. Camera changes update at most eight public geometry groups, without rewriting or uploading the resident arena. Coalesced gaps contain only conservatively rejected pages; hidden neurons remain hard barriers. If those barriers exceed the draw limit, the existing compact path preserves their visibility policy. Reports distinguish eligible, submitted and bridged indices and identify the fallback. A stable view builds one exact compact index from contiguous original runs, then draws once. Its generation key tracks exact page masks and neuron visibility, including changes hidden inside coalesced ranges. A changed view can immediately reuse the resident catalog, or keep the compact cache when its eligibility set is unchanged. Vertex buffers are shared by both index owners; memory reports include both arenas and all planner scratch buffers. This is not per-page GPU eviction or bounded initial/cache allocation/upload. Selection uses the union of eye frusta in geometry coordinates, including a conservative raster margin. Identical view stamps reuse the selection until geometry or visibility policy changes. Without the flag, the full reference submission remains available.
• Decoded mesh cache eviction uses an explicit 128 MiB byte limit. Managed CPU backing stores are deduplicated in reports. Driver GPU totals remain unknown; JavaScript heap, managed bytes and process RSS are distinct measurements.
• A separate preparation queue bounds worker concurrency while the existing network queue continues to overlap downloads. Reports retain full-session frame stalls and sample expensive memory inventories only at report intervals.
• Normal VR entry requires complete selected anatomy. Mixed diagnostic sessions require the explicit check mode; an incomplete scene cannot obtain a normal VR session merely because an empty view would meet a timing budget.
• The CPU HUD now includes frame preparation before render submission. Earlier small CPU readings excluded skeleton selection/index rebuilding and cannot establish that the workload was exclusively GPU-bound. GPU timers, when available, retain their render-query scope.

Independent references and commands

The frozen reference renderer is commit 18224c3, before page integration. The Quest capture archive contains raw colour, neuron-ID and packed-depth images, geometry hashes, camera matrices, fixed replay timestamp 20 ms, and the frozen site hash. It captures the old mixed view of 1,128 selected neurons, not a full-surface qualification. The depth/ID passes intentionally exclude context and use opaque geometry; colour captures direct rendering without the desktop composer. Raw RGBA rows start at the bottom; depth uses the pinned three.js packing shader.

uv run python scripts/check_geometry_pages.py --out runs/render-pages-oracle/new-trial

This launches an owned headless browser against local files. The reference draws original source indices independently of page selection. Six camera scenarios exercise dense all-visible geometry, close inspection, near-plane clipping, stereo projections, off-axis projection, long edges and depth ties. Colour, ID and packed-depth must match byte-for-byte. Software WebGL is a correctness test, never a headset performance result. Hashes, browser log and screenshot are retained; output cannot overwrite an existing trial.

For an explicitly selected local ?check page with completed loading:

node scripts/render_reference.mjs http://127.0.0.1:40623 TARGET_ID runs/reference/new 20
REFERENCE=runs/reference/new/reference.json CULLING=1 \
  node scripts/render_reference.mjs http://127.0.0.1:40623 TARGET_ID runs/reference/candidate 20

The second capture reuses the reference camera/projection and timestamp. Capture refuses incomplete strict anatomy, hidden pages and active immersive sessions. It restores replay/material/render-target state and writes completion metadata last. These offscreen captures do not establish XR stereo timing.

For a fixed-time desktop trial on that same fully loaded page:

PAGES=0 node scripts/observer_trial.mjs http://127.0.0.1:40623 TARGET_ID runs/trial/reference 15
PAGES=1 node scripts/observer_trial.mjs http://127.0.0.1:40623 TARGET_ID runs/trial/candidate 15

The collector restores observer state, records all intervals, visibility changes, source hashes, geometry submission and resource totals. Repeat in interleaved order; these are diagnostics, not immersive qualification. Hidden trials remain ineligible and partial content cannot pass the full-anatomy gate.

Device sampling is read-only and explicitly selects the headset:

uv run python scripts/quest_profile.py --serial QUEST_SERIAL \
  --seconds 30 --out runs/quest-profile/new
uv run python scripts/quest_xr.py --serial QUEST_SERIAL --representation skeleton \
  --pages --check --port 8895 --seconds 1800 --out runs/quest-xr/new

If using an isolated ADB server, set ANDROID_ADB_SERVER_PORT=5038 for the commands. No rules, persistent unit, clocks or safety settings are changed. The native sampler records all browser child processes and raw KGSL counters; counter reset semantics are not inferred. A bounded Perfetto capability trace was collected, but native GPU render-stage timestamps are unavailable.

node scripts/observer_gpu_probe.mjs http://127.0.0.1:40623 TARGET_ID runs/gpu-probe/new

For an actual-scheduler residency trial on a complete strict skeleton scene:

node scripts/observer_residency.mjs http://127.0.0.1:40623 TARGET_ID runs/residency/new
node scripts/observer_context.mjs http://127.0.0.1:40623 TARGET_ID runs/context/new

The residency protocol records served-source hashes before running compact / resident / resident / compact trials. Dense, anatomical-motion, close-motion and close-static views retain original geometry and shaders. Full-source colour/ID/ depth references and every measured moving colour image must match. Actual WebGL index-buffer upload calls/bytes are counted; resident steady state requires zero. Diagnostic state and the complete reference must restore identically. The context command is an explicit synthetic, non-immersive loss/recovery test; it does not simulate headset presence or certify immersive lifecycle behaviour.

This separate offscreen diagnostic remains usable when the headset suspends animation callbacks. It reads the whole framebuffer, verifies rendered and discarded image witnesses, records a clear/readback control, and restores the original shader and raster state. Altered shaders and discarded rasterization are experiments, never a faster fidelity mode. The metric includes CPU command submission, driver synchronization and readback; it is not a pure GPU timer or an XR frame rate. An earlier gl.finish-only attempt was rejected because it did not independently establish completed rendering. Readback equality for colour, ID and depth verified restoration on the physical Quest.

The recorded evidence includes the six-case raster pass and real Quest image comparisons. Culling on the same loaded scene preserves all three image buffers exactly and submits about 4.2% fewer line indices in that view. This does not establish sufficient performance. Four interleaved timing attempts produced no animation callbacks and were rejected. With explicit readback, the direct 1024-square scene took about 61–64 ms wall time; the no-raster diagnostic took about 30 ms. Removing line vertex shading naively made that diagnostic slower, not faster. Those results justify neither a shader shortcut nor a successful XR timing claim.

Resident catalog and compact-cache result

The resident evidence summary and raw archive record clean frozen source d0c4610 on the physical Quest 2. Every selected original skeleton is present: 1,128 neurons and 8,258,589 source segments.

Paired A/B/B/A medians below are serialized selection/render/readback wall times, one view, 512 square, zero MSAA, no composer. They are not XR intervals or pure GPU time. Strong timing drift remains visible in the raw samples; clocks, GC/cache effects and thermal behaviour are not controlled away.

The original compactor uploaded on average 64,640,229 bytes per anatomical-motion sample and 17,788,143 per close-motion sample. All 64 measured resident samples uploaded zero index bytes/calls. Close-static samples used the exact compact cache, not the more expensive coalesced groups. The independent arithmetic verifier passes both paired motion comparisons and the predeclared <=10% dense/close-static regression limit. This admits a component optimization only.

Four physical full-source colour/ID/depth comparisons and all 128 measured colour images matched exactly. The prior frozen 8a76f6a reference also matches all three channels at its exact camera/projection/time. Six software-raster views additionally check resident and cached indices against original indices, with nonempty ID witnesses. Synthetic non-immersive context loss/recovery preserved colour, ID, depth and geometry. None substitutes for physical immersive lifecycle testing.

Managed CPU geometry grows from 303,051,944 to 373,953,917 bytes: the extra 66,068,712-byte immutable catalog plus 4,833,261 bytes of page-run/planner metadata. Vertex attributes are shared; one compact cache is retained. GPU totals remain unknown. Warmup samples retain cache-build uploads and stalls separately: the first close-static cache build reached 68.5 ms and uploaded about 18.55 MiB. Prerequisite reference captures/initial graphics allocations are outside these timings. Initial/cache uploads, OS memory plateau and thermal qualification remain open; steady-state improvements must not hide those costs.

Two alternatives were rejected before final admission. WEBGL_multi_draw in a narrow renderer prototype preserved images but made close-static rendering roughly 45 ms versus 26 ms. Public groups without a compact cache also regressed that view. Their raw trials are retained. The final tree restores the exact upstream Three.js renderer; no vendor patch or capability override is shipped. One capability-masked fallback diagnostic failed shader compilation and has no performance result. The final frame-clock attempt had no eligible callbacks and is unscored. The dense-scene/XR performance issue remains open.

Remaining gates

The raster fixture passing does not qualify the real scene. Required next: compare the frozen real-scene images with the candidate; measure warmed, interleaved full-content trials on Quest, including all-visible and motion; refine attribution of the remaining rendering cost before changing production shaders; qualify the resident scheduler across workloads, bound initial GPU allocation/uploads and implement eviction if required by admission policy; verify memory plateau, thermal behaviour, context recovery and repeated XR enter/exit.

A mode that cannot preserve selected content within its resource and refresh budgets must remain unsupported. Do not use a fast empty/partial scene, mixed fallback, software GPU, hidden page or missing counters as qualification proof.

Browser and upload follow-up

The 2026-10-16 diagnosis adds exact-bin raster invalidation and native GPU/CPU/upload evidence. The UI optimization preserves all compared pixels without another bitmap cache. Cold offscreen reference initialization uploads about 204 MB; the dense direct loop remains about 50 ms and the existing desktop composer about 137 ms in the serialized diagnostic. Bounded initial/cache GPU scheduling and immersive qualification remain open. See the linked report for source identities, raw samples, native counter limits and the failed detailed-stage capture attempt.