Native material failure investigation — 2026-09-15

All three driving frames in Docs/QA/20260915-142737-campaign/alps-01-pass0-view0.png through view2.png were inspected. They fail the user's references: cream, blurred valley ground; white fragmented geology; repetitive forest cards. This report records diagnosed causes and source corrections, not a visual acceptance.

Pinned implementation evidence

The installed package is HDRP 17.3.0, in Library/PackageCache/com.unity.render-pipelines.high-definition@700710090fa9.

1. Runtime/Material/Lit/LitDataIndividualLayer.hlsl:208 samples every layer's base colour with ADD_ZERO_IDX(sampler_BaseColorMap). LayeredLitData.hlsl:8 explicitly documents shared samplers for textures of each type. Therefore the previous clamped orthophoto in _BaseColorMap0 also clamped all repeating forest/rock albedos to border pixels. Their nominal 5.6/7.5 m scales could not create repeating surface detail.
2. LitDataIndividualLayer.hlsl:1–79 shows that Planar uses absolute world XZ; base and detail tiling are separate when _LinkDetailsWithBase is zero. _TexWorldScale multiplies world position before the texture scale and offset. The existing geographic mapping and orientation were correct; changing the DEM or flipping the image again would introduce an error.
3. LayeredLitData.hlsl:664–688 shows main-layer inheritance adds a mean-colour correction, with low-mip samples and a special per-channel dark-value branch. It is not a plain colour blend. The former high blue compensation together with clamped corner samples and a global aerial image was an inappropriate pipeline.
4. The source aerial pixels beneath/next to the actual cameras are not summer grass. At race time 6.544 s, world (-64.52, 251.28), the central RGB is (155,150,146), and 20 m east it is (185,183,188). At 61.459 s, world (-1659.79,1128.36), 20 m west is (201,198,163). The photograph includes roofs and dry fields. Inheriting that colour across the riding corridor necessarily produces the observed pale large patches.
5. Detail packing was checked, not guessed. The meadow detail image is linear RGBA, 1024²; average R=127.50, G=127.56, B=128, A=127.37. Its red channel is correctly centred around 0.5, G/A are normal Y/X, and B is neutral smoothness. Import is linear and repeating. The aerial source is 4096², sRGB. No additional manual gamma conversion belongs in the shader colour multiplier.

Corrected material structure

• Layer 0: actual photographed meadow floor and normal/roughness, triplanar Repeat at 2.3 m. This provides the required shared Repeat albedo sampler.
• Layer 1: photographed forest litter and normal/roughness, triplanar Repeat at 3.7 m.
• Layer 2: original attributed aerial image, planar once over the unchanged DEM bounds. Only the geography's rock-weight mask selects it. Native HDRP limestone detail packing supplies independently scaled 2.8 m monochrome surface detail and normals. The brown limestone colour photograph is not recoloured blue or inherited into this layer.
• Layer 3: only the winter material uses a separate snow layer, with restrained fine normal texture. Road contact/sector physics are unchanged.

Main-layer inheritance and density are disabled. Albedo multipliers for the photographed layers are white, so the texture's sRGB decode is applied once by Unity. The root agent separately owns lighting/exposure. The importer version increased and the orthophoto's default platform size is explicitly 4096. Actual loaded dimension and shader keywords are printed by RAMPAGE_ALPINE_MATERIAL; the repeating layer-zero sampler is checked at runtime.

Forest representation

The 50 m wide, photographed 15-tree groups previously began only 86 m from the road. Their flat bases and repeating group shapes were consequently obvious at ordinary driving distances. Groups now begin at 700 m. Between the existing 86 m clearance and 700 m, each candidate produces independently grounded, irregularly spaced whole-tree silhouettes instead. This is still a billboard approximation and needs native review; it is not a claim of solved AAA vegetation.

The shared forest field and material mask now taper around local 835–1060 m plus the existing irregular boundary, approximately 1400–1625 m above sea level. This keeps conifer cards off the exposed limestone plateau. Elevation data remain byte-identical: SHA256 7f3fc34b74dc07448e973b7f9510f5a4898294fa811c4468eeff3865d6ed1414.

Verification

Offline compilation passed for 85 Runtime and 19 Editor C# sources, with zero warnings/errors. No Unity, player or GPU render was started by this subtask. A fresh native Windows capture is required to judge the corrected material and forest distribution. The earlier frames remain rejected.