Why the split is phi
You will learn
Two arguments for the phi split, and why the spec says the self-similarity one is not an optimisation.
phi_split_optimality.t27 gives two arguments and says which is which. One is self-similarity: ask that the exponent is to the mantissa as the mantissa is to both, E / M = M / (E + M). With r = E / M that reads r = 1 / (r + 1), so r^2 + r - 1 = 0 and r = (sqrt(5) - 1) / 2 = 1 / phi. The spec states that this is a defining property, not an optimisation: maximising E * M alone gives an equal split, 7 and 7 for 14 bits. The other is about rounding: round((N - 1) / phi^2) is the whole number nearest the ideal, and verify_7_7_match finds it equal to the published split for 7 of 7 formats.