[jonit-dev]

What is this PR about?

The GI calibration harness drew the cornell box darker than Unreal because it subtracted a diffuse energy loss that Unreal does not subtract in this project. This PR removes that loss, so cornell passes 10 of 12 direct-light rows instead of 4 (PRD-646).

Why does it matter?

PRD-646 is the gate for every GI verdict: a calibration scene must match Unreal before a GI change can be judged against it. Round 9 (#609) ported Unreal's 1 - E diffuse loss. Rounds 11 to 13 read Unreal's source and the level and found that Unreal compiles that loss out for this project.

• USE_ENERGY_CONSERVATION is 1 only with (SUBSTRATE_ENABLED && SUBSTRATE_GBUFFER_FORMAT>0) || LEGACY_MATERIAL_ENERGYCONSERVATION (UE 5.8.3: Engine/Shaders/Private/ShadingEnergyConservation.ush:20-31).
• The project sets r.Substrate.ProjectGBufferFormat=0 (blendable GBuffer, RenderUtils.cpp:2226-2233), and r.Material.EnergyConservation defaults to 0 (Renderer/Private/ShadingEnergyConservation.cpp:71-74).
• Measured on cornell: Unreal's specular on/off is 1 + the analytic GGX lobe (Floor 1.0274, Back 1.0138). The harness drew 1 - E + the lobe (0.9993, 1.0038). The gap equals E on each row.

Done

• ✅ Round 11: read every cornell material from the level with Unreal Python. Specular is 0.5, so F0 0.04 and F90 1, the harness's values. Nothing to port.
• ✅ Round 12: four harness arms plus an analytic D*Vis*F*NoL. The 5 cm sphere lamp is exactly the point light at roughness 1 (SinAlpha 0). The harness's 1 - E is the gap.
• ✅ Round 13: scripts/unreal-look/scaffold/render/diffuseEnergy.ts reads the project's switch (ENERGY_CONSERVATION = false), so the scale is 1 / (1 - F) and the diffuse keeps all of its light. F0/F90 come from the material. ?specular=0 now survives the term, and fails closed by name (TN_DIFFUSE_ENERGY_SPECULAR) on the owned engine, which bridges no MeshPhysicalNodeMaterial.
  • Red-green: scripts/__tests__/unreal-look-diffuse-energy.spec.ts, 2 failed before, 10 pass after.
  • pnpm unreal-look:compare --scene gi:cornell,gi:sunlit-room --layer L1, 2 runs per arm, byte-identical within each arm:

Arm	cornell	sunlit-room	worst dP50	worst deltaE
BEFORE	4 of 12	40 of 40	0.059	1.10
AFTER	10 of 12	39 of 40	0.029	0.34

• Visual judge (fresh Opus, BEFORE/AFTER/UNREAL sheets of all 8 poses): NEUTRAL on every sheet. No artefact, no hue or shadow change. The 1 to 6% linear gain is below what the sheets show, and the compare rows carry it.

No file under packages/ or templates/ changes.

Still open

• ⚠️ AC-2 stays unticked. Three rows miss: cornell Box_Tall in two poses (2.9% and 3.2% low) and sunlit-room sun15-window Floor (3.1% bright, dP50 0.011 against 0.010; it was 1.4% bright and passing before). No constant was added to win them back.
• ⚠️ Engine follow-up: a MeshPhysicalNodeMaterial catalogue entry would let ?specular=0 run on the owned engine.

Checks

• pnpm exec vitest run scripts/__tests__/unreal-look: 47 files, 678 tests pass.
• pnpm check:docs passes. unreal:status is up to date. Biome is clean on the changed files.
• pnpm prd:progress: 3/4 phases, 11/12 boxes, 2/3 ACs, prd:75%.