Performance rules

Performance optimization must preserve the accepted scientific implementation.

Rules:

• measure before optimizing;

• keep science acceptance separate from runtime acceptance;

• use the same model/input/data/compiler when comparing timings;

• record CPU features and thread settings;

• prefer generic source-equivalent optimizations over element/model-specific shortcuts;

• do not reopen frozen Type50, trajectory, matrix, or transport semantics merely for speed;

• after a behavior-affecting optimization, qualify one fast model before broadening coverage.

Historical performance reports remain available through version-control history and release tags. They are useful attribution evidence, not current user-facing performance guarantees.