# Classified distant-point raster check

This bounded check addresses only `original-data-37/frame-fade-d50-forced.png`. That capture retains the actual recorded system address207931040981 and position(-388.375,-369.5,25494.34375), but uses explicitly controlled camera, membership and appearance inputs. It is not a live original-game capture.

The 1878×1148 image centres the point on an integer pixel coordinate. The camera focus is64ly behind the point and its distance is800ly, making point depth864ly. The recovered Core lens gives physical focal length988.190007525405px. The original52-byte producer quantises `.7` to half `.7001953125`, so the projected quad half-extent is `.8008402906581346px`.

The original `BackgroundStars` VS multiplies the six actual cornerZW ±1 values by `TEXCOORD3.z`, offsets world position with shared camera Right/Up, and applies `cb0[4..7]`. It contains no pixel-centre rounding, half-pixel translation, or minimum-size term. Production projects the same quantised world size to screen space; viewport coordinates are logical pixels, while its raster viewport is scaled by DPI. Scaling both projected centre/radius and viewport consistently preserves physical pixel geometry.

The key loss occurs in the original PS, not missing quad coverage. For the proved default material `(pInternalGlowX,Y,Z)=(2.75,.17,1.903)`, before texture and colour multiplication the radial factor is:

`2.75 + 1 - radius / (size * .17)`.

Since `Z>1`, the PS's additional `light*(1-saturate(Z))` term is zero. RGB becomes nonpositive outside radius `.6375*halfExtent = .5105356852945608px`, then original `max(...,0)` clamps it to zero. At an integer-centred point the nearest single-sample D3D11 pixel centres are at distance `sqrt(.5²+.5²)=.7071067811865476px`. The quad therefore covers four samples but none is inside the positive-RGB region. The original PS still writes instance alpha1.

## Actual DXBC oracle

`research/GalaxyStellarRasterProbe` runs only owned offscreen D3D11 WARP targets. It uses the hash-verified original VS (`4bfe9d...5b09d`) **and** PS (`a9eef9...1d9c`), the original half4 UV/corner order, actual52-byte per-instance input layout, original BC1 array and linear sampler, and the same camera/material inputs. It separately executes production `VSStar/PSStar`. It does not use a bridge VS in the original path. The successful frozen App Core DLL is referenced directly to avoid rebuilding unrelated concurrent source edits; its hash is recorded.

The R32 float16×16 patches give:

| Physical pixel phase | Original RGB samples | Production RGB samples | Original/production alpha samples |
| --- | --- | --- | --- |
| (0,0) | 0 | 0 | 4/4 |
| (.25,.25) | 1 | 1 | 4/4 |
| (.5,.5) | 1 | 1 | 1/1 |
| (0,.5) | 2 | 2 | 2/2 |

All four cases are repeated at DPI1,1.5,2. Integer-centred patches have exactly zero full RGBA difference. Other phases have identical coverage, with maximum RGBA vector difference `.0028468987` from original world-space float interpolation versus the production screen-space computation. No minimum pixel size or extra half-pixel offset is added. Full input hashes and raw float patches are in `artifacts/galaxy-stellar-raster-validation/report.json`.

Reproduce with `dotnet run --project research/GalaxyStellarRasterProbe --artifacts-path artifacts/galaxy-stellar-raster-build`. Only this independent probe is built; it does not build the App or invoke Native/game code.

## Boundaries

This establishes the observed black RGB under the matched **default material, unjittered projection and single-sample** state. It does not establish which live alternate-glow branch the original game selects, the complete D50 row10 consumer, temporal projection jitter, original runtime multisampling/antialiasing or final scene presentation. Those can change sample phase or the radial term and require their own evidence. A retained instance/hit counter is not a visible-RGB count, and this result is not a global galaxy-map visual pass.