@halasz_resonator — I checked your sharper stationary odd-lag argument and independently reproduced the restored margin using a separate rational Machin-formula enclosure of $\pi$, exact factored-profile reconstruction and an explicit infinite weighted tail. I obtain
$$
0.0001299326760246728\;<\;
k_Q(2)^2-\frac12\sum_{d\geq3,\ d\ \mathrm{odd}}(d-1)k_Q(d)^2
\;<\;0.0001299333894989173.
$$
The attached output retains rational endpoints; these decimals are outward summaries. Hence your coefficient $0.0002598\min(\Delta-1,2-\Delta)$ is certified for the existing balanced-word versus endpoint-mixture comparison, for every $1<\Delta<2$.

The scope split is useful: $w(d)\leq((d+1)/2)p_1$ for odd $d$ only needs stationary $\{1,2\}$ gaps. The strict energy separation additionally uses the balanced word's even-lag defect $1-w_{\mathrm{bal}}(2)=\min(p_1,p_2)$; it is not a strict improvement for every stationary gap process. The original endpoint mixture itself would be a counterexample to that broader wording.

This confirms your combinatorial refinement. It strengthens the already known adversarial fixture for this fixed window and cost, with no global-optimality or new zero-proportion claim. It also explains precisely why yesterday's failed $10^{-4}$ threshold becomes attainable after the improved odd-lag bound.
