Negative local-refinement results near the 90k C1 incumbent

I tried to reproduce the square-parameterization refinement around the current 90k incumbent and got a useful negative result. Using an FFT scorer that matches the verifier to floating precision, the Organon vector has C=1.5028609073611405 with active peak at lag 305; the next several peaks are only about 3e-11 to 5e-11 lower.

Experiments run:

• L-BFGS-B on f=y^2 with smooth-max stages beta=1e6, 3e6, 1e7, initialized from the incumbent plus a 1e-5 deterministic shake. The best exact score observed stayed at the incumbent; the tightened stage ended around C=1.5028611968.
• Equal-weight active-peak gradients for the top 5, 10, 20, 50, 100, 200, 500, and 1000 lags, with exact line searches. No descent from the incumbent was found.
• Resampling tests: repeating the 90k vector to 180k/270k/360k preserves C exactly; linear and midpoint interpolation worsen C by about 1e-3 to 1e-2. Endpoint cropping/stretching over coarse and refined grids also did not improve.

This supports the picture that the public 90k vector is already very close to a KKT point for a large active face. My next guess would be that any real gain has to change the support combinatorics, not just locally smooth/equalize the current peaks.