Optimize only a measured route-phase bottleneck. Safety semantics and source ordering must remain identical.
- Reuse one prepared hazard window inside a phase decision when multiple exact queries share the same snapshot.
- Keep decoded ECL/enemy state scalar inside frame/entity loops; this reduced a retained Stage 3 local decision from roughly 25.6 ms to 9.5 ms.
- Advance temporal hazards once and use spatial pruning only before the same exact collision predicate.
- Minimize physical counterexamples before retaining them.
- Deduplicate physically aliased reachable states, especially at movement clamps.
- Move static timeline/ECL decoding offline.
- Precompute phase policy structure and source event schedules.
- Keep online conditioning proportional to the actual branch state.
- If a phase helper is measured too slow, move that small exact helper to native code and parity-test it.
- Reject an optimization that wins by stale projections, omitted source events, partial results, or changed collision semantics.
Do not resurrect a global adaptive horizon ladder merely as an optimization. The owning route phase selects the work it needs; the runtime enforces its deadline and falls back only to a freshly Hard-certified held action or an explicit authority stop.