docs: session-154c - viewport/clip/RT0 closure with driver semantics and the shader-crutch finding: the 144-byte vs_hex/ps_hex test image overwrote the client SASS and writes no position; removing it lets real SASS survive but stores remain zero, so the next checks are the VS position/attribute fetch and the client upload path

1 files changed, 24 insertions, 0 deletions

diff --git a/docs/04-implementation/04-real-render-walls.md b/docs/04-implementation/04-real-render-walls.md
index 4500f5d..3744891 100644
--- a/docs/04-implementation/04-real-render-walls.md
+++ b/docs/04-implementation/04-real-render-walls.md
@@ -1006,3 +1006,27 @@ replayed context a missing viewport clips all geometry, which matches the
 zero-fragment observations. The next experiment synthesizes a gated viewport
 prelude at the window preamble and re-runs the bracket; a pass is any changed
 page inside the render-target range.
+
+The synthesized viewport was then closed with the driver's own semantics: the
+live viewport macro is MME slot 33 (CALL method 0x3908), the clip rectangle is
+0x0c00/0x0c04, and the clip control must read 0x00021800 (Z_CLIP_RANGE
+ZERO_ONE | GEOMETRY_CLIP FRUSTUM_XYZ) - the earlier value 1 left the geometry
+clip in its WZERO mode. The colour-target closure was also found: the client's
+pixel shader has MRT_ENABLE=0 and writes target 0, while the window binds only
+target 1. Both a corrected viewport/clip prelude and a target-0 bind with a
+two-target SET_CT_SELECT are now gated behind module parameters, and the
+pair-state orphan END fix stopped the EC0x0d poison after held windows. With
+all of that in place and the DRAW macro executing (DRAW_VERTEX_ARRAY, 60/46
+vertices, QUAD_STRIP), the whole-shadow bracket still reports no render-target
+page change.
+
+The remaining input is the shader content. The program pages carried valid SPH
+headers but their code was a 144-byte SPH+EXIT test image supplied through the
+`vs_hex`/`ps_hex` debug parameters, which overwrote the client's real SASS;
+that image writes no position, so every vertex is degenerate. Removing the
+test images lets the real SASS survive at the bound program addresses, but the
+engine still stores nothing. The open checks are: whether the real vertex
+shader writes position and reads the attribute stream; why the client's own
+upload accounting still reports zero applied bytes; whether the vertex fetch
+bases hold the geometry the guest writes; and a positive control that draws a
+constant-position full-screen triangle with a colour pixel shader.