Revert MapOutputRectToInputRects sub-rect propagation

Earlier today (1ea4365) I changed CustomPixelShaderEffect::Map-
OutputRectToInputRects to honor the queried output sub-rect, hoping it
would unlock leaf-side preview-resolution scaling and per-input-rect
routing for Split. The Split-pivot regression check passed at full-
canvas preview, but the broader sub-rect propagation breaks thumbnail-
sized preview panes: D2D queries a small output sub-rect to fill the
thumbnail, our chain dutifully renders only that sub-rect, and the
preview pane shows a small rendered region in the upper-left of an
otherwise-black canvas. User-reported via Gamut Highlight thumbnail.

Going back to the safe baseline -- always return m_lastOutputRect from
MapOutputRectToInputRects regardless of queried sub-rect. This guarantees
the intermediate textures match the full effect output rect (the with-
TRIVIAL_SAMPLING contract) and keeps thumbnail / preview-pane rendering
correct.

Trade-off: D2Ds lazy-eval cant propagate sub-rect demand through our
chain, so leaf-side approaches to preview-resolution scaling cant
rely on this mechanism. The Scale catalog node ships the explicit-
insertion model instead -- which is what we landed on anyway after the
B/C architecture options were abandoned.

Tier 2 (per-input-rect routing for Split) would still need this
mechanism but it was already deferred indefinitely.

154/154 tests pass.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: Spruill-1

Effects/CustomPixelShaderEffect.cpp

@@ -252,31 +252,28 @@ namespace ShaderLab::Effects
     }
 
     IFACEMETHODIMP CustomPixelShaderEffect::MapOutputRectToInputRects(
-        const D2D1_RECT_L* outputRect,
+        const D2D1_RECT_L* /*outputRect*/,
         D2D1_RECT_L* inputRects,
         UINT32 inputRectCount) const
     {
-        // Pass the queried output sub-rect through as the input demand. With
-        // D2D1_PIXEL_OPTIONS_TRIVIAL_SAMPLING set, our shaders read input
-        // pixel (x,y) for output pixel (x,y) -- so the input rect needed is
-        // exactly the output rect being requested.
+        // Always return the FULL output rect as the input demand, ignoring
+        // the queried sub-rect. This is the safe-with-TRIVIAL_SAMPLING
+        // baseline: it ensures the intermediate textures match the full
+        // effect output rect so TEXCOORD maps 1:1 with GetDimensions()-
+        // normalized math, and so thumbnail / sub-rect previews show the
+        // *whole* effect output rather than just the queried region.
         //
-        // Honoring the sub-rect (rather than returning m_lastOutputRect for
-        // every input) is the linchpin of D2Ds lazy-evaluation propagation:
-        // if a downstream consumer (e.g. the swap-chain present at a small
-        // canvas size) requests only a sub-rect of our output, that demand
-        // cascades upward and every upstream effect renders only what's
-        // actually needed. For preview-resolution scaling on heavy 4K HDR
-        // graphs this can reduce upstream pixel-shader work by 4x or more.
-        //
-        // Split Comparison and other effects that need the full output rect
-        // for math (pivot recentering, etc.) read the full dimensions from
-        // host-injected OutputW/OutputH cbuffer fields, not from the per-
-        // dispatch rect -- so this change is transparent to them.
-        const D2D1_RECT_L safe = outputRect ? *outputRect : m_lastOutputRect;
+        // An earlier attempt at honoring the queried sub-rect to enable
+        // D2Ds lazy-eval propagation (for preview-resolution scaling and
+        // per-input-rect routing) caused the symptom that thumbnail-
+        // sized preview panes rendered only a small sub-rect of the
+        // chain into the upper-left of the canvas. Reverted; preview-
+        // resolution scaling is shipped as the explicit Scale catalog
+        // node instead, which constrains the source-side rect rather
+        // than relying on leaf-side sub-rect demand propagation.
         for (UINT32 i = 0; i < inputRectCount; ++i)
         {
-            inputRects[i] = safe;
+            inputRects[i] = m_lastOutputRect;
         }
         return S_OK;
     }