Train: Optimised single_pass — soft-inverse τ30 + L2 + sky0.1 area-pooled + 10% zero-SNR (C3) 914fa85c

Supersedes config 96 (C2) for a cold-start soft-inverse-depth experiment on model “SD 2.1. Mari”. Primary changes: 1. Uses model.depth_mode=metric_soft_inverse with one global tau=30 m. The forward mapping is y=(d−tau)/(d+tau), so near depth is low and far depth approaches +1. 2. Uses a single L2 diffusion-prediction loss from epoch 0; removes the C2 L2/Charbonnier ramp. 3. Samples the zero-terminal-SNR endpoint for 10% of examples and samples the remaining 90% uniformly from first through penultimate timestep. 4. Supervises finite sky targets at 700 m with weight 0.1 and area-pools depth supervision weights. 5. Sets the model decode/evaluation/export cap to 700 m, the shared dataset validity minimum to 2.5 m, and the general validity maximum to 700 m. Soft inverse intentionally has no normalization_min_depth_m; 2.5 m is only a dataset validity cutoff. 6. Enables explicit mask-aware depth resizing. VKITTI2 remains a necessary source-specific exception at valid_max_depth_m=650 m because its 16-bit-centimetre sky sentinel is 655.35 m. Raising that validity threshold to 700 m without a semantic sky mask would classify the sentinel as ordinary valid depth and defeat sky supervision. The finite sky target and model cap remain 700 m. RDS and MUSES use the shared 2.5–700 m validity interval. Euler’s RDS metadata reports a raw 0–1000 m range, so RDS samples above 700 m intentionally enter the far/sky fallback and receive the finite 700 m target. This successor does not resume a metric-log checkpoint; the inherited launch script starts train_joint.py without --resume, avoiding representation reinterpretation.