Direction 28: Nothing is final until it is measured
(Robert, session 2) Every decision here — especially the performance-shaped ones — stands only until a benchmark, eval, or test says otherwise. Compile-time performance is of paramount importance, and runtime speed close behind, but both are measurable, so neither is settled by argument. Consequences: the project carries a benchmark suite and an eval suite from early on; every performance claim in this journal is a hypothesis with a named way to check it; a decision that measurement contradicts gets reopened without ceremony. This applies to the overflow-check cost, the interpreter-vs-native split, the incremental-rebuild targets, the C-via-Zig backend choice, and any law number.