bartulem/usv-playpen: v0.11.9

Torus geodesic MAEs now reported by the univariate manifold screen

The two analysis-only torus distance columns -- density_geodesic_mae (the density-ratio graph geodesic) and pullback_geodesic_mae (the decoder-Jacobian pullback geodesic) -- were computed only in the acoustic-manifold selection stage. The univariate screen that precedes and feeds selection therefore reported a strictly smaller torus metric set than the stage it screens for.

run_univariate_training now emits both columns too, mirroring selection: per fold across all three strategies (actual / null / null_model_free) and in the honest held-out refit. The geodesic reference map is fold- and feature-independent, so it is built once per run and cached on the runner rather than rebuilt for every feature (the pullback-Jacobian build is the expensive part, and the univariate runner runs once per feature). The decoder is resolved through the same configure_path chain as selection, so the frozen QLVM decoder .npz is reachable from the compute cluster. Torus-only; every column degrades to NaN (and never aborts) when geodesic metrics are disabled, the metric is not 'torus', or the decoder is unavailable -- only the pullback column degrades in that last case.

The von Mises selection criteria (macro vm_logscore / micro vm_logscore_pooled) were already reported by the univariate screen, so this closes the remaining metric-parity gap between the screen and the selection it feeds.

Re-run note: existing univariate manifold results predate these two columns; re-run the univariate screen on a 'torus' metric to populate them. No selection decision changes -- the geodesic MAEs are diagnostic columns, not the screening / selection objective (which remains vm_logscore).