Bounded Alps02 cutting pilot — native geometry verified, art goal open

Source evidence: Docs/QA/20260916-013504-campaign/alps-02-pass0-view1.png, route progress 1978.616 m, bike (-1378.781, 459.240, 2724.155). The broad pale wall is the graded terrain/road-cut overlay, not a new floating rock card. The original apron has lateral rows at 6/8/14/27 m. At this pose its outer 14–27 m span rises about 23.1 degrees while the old rock mask starts at 32 degrees, producing a straight green toe against the exposed wall. That numeric cross-section ignores local banking and is a diagnosis estimate, not collision evidence.

AlpineCutVisual applies only to alps-02, route 1750–2200 m. It is a visual pilot with a 24 m end fade and unchanged inner 14 m corridor. Uphill terrain is selected using the unchanged DEM relative to road height. Shared world-XZ slope/material sampling blends the apron and nearby terrain; the local talus blend starts at 18±3 degrees and reaches full rock at 42 degrees. Outside this bounded envelope the existing material rule remains.

The actual apron through 27 m is not displaced. Beyond 27.5 m, rising smoothly to full strength at 34 m, the rendering copy receives deterministic world-coordinate relief of 0…0.35 m. Original mesh perimeters and modified/unmodified quad boundaries are pinned with a 2 m fade. Existing tree anchors can be embedded by at most 0.35 m; positive-only relief never lowers the ground beneath them. No route, terrain source, material-map bitmap, SceneryGroundHeight, Rigidbody, or collider mesh is edited.

Subdivision is at most 8×8 per original quad, aiming for 2.5 m samples. Cells crossing the original quad diagonal are clipped into its original two planes, preserving the complete surface's 0…0.35 m displacement bound even on warped quads. Base surface and rock overlay use the identical resulting vertex array and normals. Both render passes count toward a hard additional budget of 59,998 triangles; rejected quads retain their original geometry and their modified neighbours pin the shared boundary.

Construction is bounded to at most four intersecting 512 m terrain tiles plus apron chunks in this route segment. Each boundary query uses a spatial 4 m bin, not the full tile perimeter. Shared normals and exact ground samples are cached by XZ during construction and all construction caches are cleared and trimmed in Finish. There is no per-frame callback or persistent static stage cache. The native run below confirms nine affected meshes, 9,526 cached normals, 46,297 ground samples and 20,925 local boundary-edge checks. Construction wall time was not separately instrumented.

Validation so far:

• Managed compilation: 98 runtime + 22 editor files, zero warnings/errors; no Unity process started by the author.
• ArtSource/Alps/audit_cut_visual_topology.py checks all 64 subdivision combinations: 3,296 triangles have positive winding, stay on the corresponding original plane and cover the unit quad with maximum error 1.34e-15. Worst individual quad is 276 triangles including both passes, within its conservative reservation. Report: Docs/QA/alpine-cut-visual-topology.json.
• New script metadata has a valid 32-hex GUID and actual line breaks.
• At native construction, each affected original collision mesh is SHA-256 fingerprinted from vertex float bytes plus triangle indices before and after visual assignment. The collider mesh reference must remain identical. The recorded native run confirms this for all nine affected meshes. This verifies the checked geometry and sharedMesh reference, not every serialized physics setting or human collision feel.

Native run, 16 September 2026, 02:22 Vienna

Evidence: Docs/QA/20260916-022247-campaign/player.log lines 70–80, invocation.json, campaign-report.json; extracted audit Docs/QA/alpine-cut-visual-native-audit.json. Windows assembly SHA-256: 91D6BB40B36C858FBB59826AEFFFAAC8F707F94ED82FE4975404598552C8809F. Unity 6000.3.24f1, Direct3D11, Intel UHD Graphics, exit 0; both solo AI/Rigidbody stages passed, with zero crashes, recoveries or missed-checkpoint frames. This is separate from human/controller and deliberate collision testing.

The pilot added 26,267 render triangles, including ground and overlay, below the 59,998 cap. Maximum measured relief was 0.3499995 m. All five apron meshes logged exactly zero displacement; all nine original collision meshes retained their hash and reference.

The structured audit distinguishes the native assembly hash, raw source-file hashes, and the managed compiler's UTF-8/LF-normalized source hashes. Both pilot source files still match the normalized hashes in the successful 00:15:32 UTC compile manifest; their raw and normalized hashes are also identical in this snapshot. The source hashes are not the native assembly hash. The topology report remains explicitly a CPU proof, now accompanied by native mesh evidence.

Visual review of alps-02-pass0-view1.png: the previous sharp green toe now has a softer, mottled talus transition. This is an incremental improvement. The large pale, planar rock wall still looks manufactured and does not meet the photorealistic/AAA art target. Keep this bounded pilot without extending its scope. No physics geometry was changed; no human collision-gameplay acceptance is claimed. The observed Intel Alps02 average of 20.03 FPS is also not a performance acceptance for the intended gaming GPU.