RuLab GPU realism: photographic surfaces, HDR, quality controls and evidence

[ruvnet]

RuLab now has a browser GPU graphics upgrade that improves surfaces, illumination and reflections while preserving Gaussian imports, capture playback and deterministic WorldGraph replay.

Implemented:

1. Seven bundled CC0 assets with SHA256 provenance, totaling 3.31 MiB.
2. Photographic PBR maps, HDR illumination and exterior, FXAA, optional bloom and mesh floor reflections.
3. Quality and exposure controls with renderer identity, resolution and mode in exported evidence. Drawing buffers are capped at 1 million pixels for Performance, 2 million for Adaptive and 4 million for Quality.
4. A release harness that rejects missing or altered photographic assets.
5. npm run benchmark:gpu for a graphical desktop. It rejects known software renderers, exercises four views over 60 seconds, and requires 300 samples with p95 frame time at most 33.3 ms.

Verified result:
All seven fixed validation gates passed with zero skipped gates: 112 unit tests, 23 harness tests, 88 Rust tests including doctests, TypeScript, WASM build, production build and all 16 desktop/mobile software WebGL tests. Standard repository CI also passed. Desktop and mobile Quality screenshots were inspected.

CI: https://github.com/ruvnet/worldgraph/actions/runs/34071798432
Standard CI: https://github.com/ruvnet/worldgraph/actions/runs/34071798469
Head: 70ac6bd
Tested merge: 42aa02e
Identical source tree: a55d1d745bd2b4dfcd7372391d4ba12669caf1f7
Evidence artifact: 10000912483, SHA256 1caf2e3f306de8905600244fc3f6b6d132dd340c03ea9caa9b3215fdc958877a
Build artifact: 10000912843, SHA256 1168e7ba025d55fdf46111198293a45d980012eca6e147e3387a3b7bf946bf5a
All seven log hashes and 18 build asset hashes were independently compared with the downloaded CI artifacts. MCP evidence validation also returned consistent and complete, while explicitly disclaiming independent execution attestation.

The first run passed 14 browser cases and failed two because the quality selector lacked an exact accessible label. That was fixed, verified in the browser, and the complete suite passed on the final tree. No failure was waived. Superseded PR runs now cancel automatically, while main release serialization remains unchanged.

Limits:
Physical GPU benchmarking has not run here; CI uses SwiftShader. Geometry remains an authored concept, not a reconstructed facility or learned world model. Reflections include meshes and exclude Gaussians from the secondary camera. Imported captures retain baked appearance and do not inherit the authored floor reflection. Passing software WebGL tests does not establish photorealism or hardware speed.

ADR 207 and README cover setup, controls, provenance, rollback and hardware acceptance. Publication hold is recorded in https://github.com/ruvnet/core-memory/issues/61.

DO NOT MERGE OR DEPLOY. The user requested that the current published Explorer remain unchanged. No Site update, main merge or Pages deployment was performed. The CI deploy job was skipped.

[ruvnet]

Dream exact-head review — INCONCLUSIVE.

The software envelope is strong: the reported unit, harness, Rust/doctest, TypeScript, WASM, build, and 16 software-WebGL cases agree with the two green exact-head workflows. Asset hashes and the publication hold are explicit. The frozen physical-performance hypothesis is nevertheless untested because CI uses a software renderer and benchmark:gpu did not run on target hardware.

Keep the PR draft and the publication hold. Acceptance requires the declared 60-second/300-sample hardware run on each target class, p95 frame time ≤33.3 ms, renderer/driver/device identity, context-loss count, resolution/quality mode, captured artifact hashes, and adversarial altered-asset failure. Supply-chain hardening is now tracked separately in #11; it should preserve the deploy authorization boundary.