Prove emergent divergence with a lexical-vs-state A/B control sim

When replacing a lexical (word-overlap Jaccard) similarity function with a state/property-based signature, keep the old lexical version as a *cold-start fallback* that only fires when a node has no state yet — this keeps every existing test byte-identical while new behavior only activates once state accrues. To *prove* the new function actually causes structural divergence (not just claim it), build one sim that runs N worlds sharing identical vocabulary but differing only in the underlying state, and run the SAME construction twice: once in state-mode (expect divergence, measured as edge-set symmetric-difference > 0) and once with signatures stripped so it falls back to lexical (expect edit-distance 0). The A/B on one code path isolates spelling as the cause, far more convincing than a single-mode assertion. Also assert bit-for-bit reproducibility from a fixed seed.