Status: accepted (2026-08-02)
Context
v1 stored only leaf states and recomputed ancestors constantly - four separate code paths existed just to fetch ancestors, plus a 376-line precomputed cache module (descendant sets, ancestor paths, a pairwise LCCA matrix). The Appendix D algorithm assumes the full configuration is available, so v1 also had to reconstruct entry sets from leaves.
Decision
The active configuration stores all active states, ancestors included, as the spec assumes; "leaf states" is a derived view. The Machine compiler interns state IDs to integers and lays states out in a flat array in document order with parent pointers and descendant index ranges. Ancestor/descendant tests, LCCA, document-order sorting, and exit-set computation are integer/range comparisons on that layout.
Consequences
- The Appendix D port maps directly onto the storage model (ADR-0002).
- The hierarchy cache module and the four ancestor paths disappear.
- Configurations are
MapSets of integers; string IDs appear only at the API boundary (parsing in, event/introspection out). - Slight cost: a translation step at the boundary, and debugging views must map indexes back to IDs (the Machine keeps both directions).