D112 - The GPU is driven by submissions

Status: decided Date: 2026-09-26

orbistoun-gpu::pipeline takes a submitted command buffer and a one-method reader of guest memory, and derives everything - shaders, addresses, stages - from what the guest wrote. A shader is read as a window up to its end-of-program instruction; a window that runs out is refused, not truncated. Shaders the guest registered by name and shaders inferred from register writes both run, and every disagreement is reported. A shader that will not translate is reported with its address and reason, never skipped.

Why: this hardware generation has no high-level graphics call to intercept. A truncated shader is a genuine prefix that looks correct. Agreement between the two routes is the only evidence the register table is right.

Rejected:

• A translator handed shaders by the emulator: nothing would ever hand it one.
• Stopping at the first route that answers: discards the only check on the register table.
• Skipping an untranslatable shader silently: a frame that drew nothing, unexplained.