D014 - Throughput is an explicit dimension of completeness for oops-gl

Status: decided Date: 2026-09-26

Completeness for oops-gl includes throughput alongside API conformance. A pipeline that renders a correct frame but runs at unplayable frame rates on real geometry is incomplete.

Why: Under conformance-only gates (D008), GL 1.4/1.5 and GL 2.0 probe suites passed, yet Craft ran at 0.05 fps (150 µs per triangle) because GL 2.0 vertex shaders were interpreted on the CPU in glsl_exec.c and batched geometry was broken into individual 3-vertex draw packets. Compiling vertex shaders to native gfx1030 machine code (glsl_vs.c) and adding a resident vertex buffer draw path in gl_draw.c brings triangle submission to the GPU while retaining glsl_exec.c as the reference interpreter.

Rejected:

• CPU-only vertex pipeline optimization: Cannot bridge a 200x gap when 89% of the frame budget is CPU vertex transform.
• Full GPU shader compilation without resident buffers: Interpretive or per-triangle packet emission still bottlenecks on command processor and bus throughput.
• Replacing glsl_exec.c: The CPU interpreter is the golden reference for conformance and debug oracle validation.