10. Instruction semantics are separate from timing

Context. The two backends each contained parts of execute: the in-order backend executed vector instructions against the architectural registers itself, decoding lived in the frontend, and the retire effects lived in commit. Each backend's copy could drift from the other's, and a semantic bug fixed in one stayed in the other.

Decision. The exec layer defines what every instruction does, once: decoding into an exec::Inst with its control signals, executing it, the values loads produce, vsetvl, the vector and floating-point units' arithmetic, and the effects commit applies on retirement. Execute reads the hart through the ArchState trait (the hart, CSR reads, and whether to trace) rather than through a pipeline's view, so it depends on nothing in uarch. The pipelines decide when an instruction may issue, which unit holds it for how long, and when its result becomes visible; for what the result is, they call exec.

Consequences. Both backends compute identical results by construction, so a difference between them is a timing-model bug. The riscv-tests and the spike vector cross-check exercise one implementation of each instruction. exec sits below the simulation kernel in the layering and can be tested without building a pipeline.