[lohedges]

Fixes the calibration of the adaptive replica exchange convergence checks.

Offline replays of three 5 ns GSK3B ABFE bound legs showed that the checks flag converged window pairs far more often than intended:

• The two-halves drift check estimated its error linearly, ignoring how strongly the exponential average depends on a few extreme samples. On block-shuffled data, which has no drift, it fired 25–33% of the time instead of ~5%, and up to 100% for pairs with wide energy distributions. It now uses a jackknife over contiguous batches, which gives 5–9%.
• The forward/reverse check (|EXP(fwd) + EXP(rev)| > 1 kT) fires on perfectly converged data once the spread of the energy differences exceeds ~2 kT, because of the bias of the exponential average. It is now only applied when both spreads are at most 2 kT. Wider pairs are still covered by the variance ranking and the swap acceptance check.

In the replays, about 30% fewer pairs are flagged per ranking, so more windows can go cold. The test helper's hard pair is now consistent, so the ranking tests no longer get their hot windows from a flag.

[lohedges]

CI failures are because of this issue. Should hopefully be sorted soon.