Verify a perf cache against the uncached path with normalized equality

When replacing a per-call full-table scan with a once-per-tick prebuilt cache (e.g. an adjacency map from one edge scan instead of scanning edges per node), add a regression test asserting the cached path produces EXACTLY the same result as the scan path — normalize first (sort neighbor lists, round floats) since dict/list ordering differs. Also: gate the new behavior behind an opt-in meta flag so existing worlds are byte-for-byte unchanged, and assert the flag-off path is a literal no-op. Debugging lesson: an apparent bug (a merchant shipping to a NON-adjacent node) turned out to be correct emergence — an earlier tick pass had laid new trail edges, so the node WAS adjacent at decision time. Before assuming a bug in a multi-pass simulation tick, replay the exact pass order and inspect state between passes; standalone reproduction that omits earlier passes can hide the real cause.