fix(calibration): score the differentiable loss on the calibration period

[DarriEy]

Problem

kge_loss and nse_loss dropped warmup and then scored everything that remained. With a calibration/evaluation split configured, that span covers both windows.

Two consequences:

1. Evaluation-period leakage. Gradient-based optimizers were trained on the held-out evaluation period, so the split-sample test was not honest.
2. Mismatched reporting. The score the optimizer reported was computed over a different window than the Calib_* metrics written at final evaluation.

HEC-HMS had no completed ADAM run in the P3 ensemble, so this is fixed before it could produce misleading numbers. First run on fixed code: best 0.9010 == Calib_KGE 0.9010, with Eval_KGE 0.7946 genuinely held out.

Every other algorithm in the ensemble agreed exactly, because they score through _run_simulation rather than the JAX loss. L-BFGS was unaffected only because its configs request finite-difference gradients.

Fix

Thread a calibration slice through both losses, both backends, and the gradient-function factories. The worker supplies it from the shared InMemoryModelWorker.get_calibration_slice(), so all packages share one implementation. Requires the companion SYMFLUENCE PR.

Tests

New test_calibration_window.py (6 tests): a slice covering the whole post-warmup record is a no-op, a narrower slice changes the score, and two disjoint windows do not collapse to the same number.

These were mutation-tested — disabling the slice in _eval_window fails 4 of the 6 — so an accidental revert cannot pass silently.

Same defect found and fixed across jHBV, jTOPMODEL, jSAC-SMA, jXAJ, jHEC-HMS, jFUSE and cFUSE.

🤖 Generated with Claude Code