Train: MultiSet + Muses Val V2 + Sky-SDF 16ch + C-RADIOv4 ece594b5

Copy of "Train: MultiSet + Muses Val V2 + Sky-SDF 16ch" (config 111, v5) with the frozen C-RADIOv4-H feature-conditioning pathway enabled, and nothing else changed. WHAT IS NEW. model.feature_conditioning injects one frozen, spatially aligned C-RADIOv4-H feature map into the SHARED encoder, by adding a separate zero-initialised conv_f to conv_in's output. A convolution over a channel concatenation is the sum of convolutions over the parts, so this is algebraically the concatenation -- while leaving conv_in's 16-channel width and every pre-feature checkpoint untouched, and being exactly identity at step 0. Both task decoders and the sky decoder see it through the trunk. The intent is a fifth cross-task coupling channel: all four existing channels couple the two predictions to EACH OTHER, which two branches can satisfy by hallucinating the same wrong scene; a shared external anchor makes agreement have to be agreement with the anchor. See docs/feature_conditioning.md and docs/plans/foundation_feature_conditioning.md. DELTAS FROM CONFIG 111. (1) model.feature_conditioning.* added, enabled=true, backbone=c-radio_v4-h, channels=32, autocast_dtype=bfloat16. (2) training.feature_adapter_learning_rate=6e-5 -- REQUIRED when the pathway is on; without it the zero-init conv_f rides in the backbone group and train_joint.py fails fast. (3) training.weight_decay_groups.feature_adapter=0. (4) run.sh exports TORCH_HOME and aborts at startup if the torch.hub cache is cold, because Euler compute nodes cannot download the backbone. (5) segmentation_detach_encoder_features and use_deterministic_vae were unset in config 111 and had to be typed at every launch; pinned here to the values run 2026-09-11_16-19-37_e3fc actually used (both true). UNMEASURED. channels=32 and feature_adapter_learning_rate=6e-5 are reasoned defaults, not optima. No GPU has ever executed this pathway; throughput and peak memory are unmeasured, and one ViT-H/16 forward per image per step at batch_size 24 is a real cost. gpu_type still reads rtx_4090:1 from the source config -- set the A100 at launch. WATCH IN THE FIRST HOUR. rgb.train.stat.feature_contribution_ratio must rise off exactly 0 (conv_f starts at zero, so a flat 0 means the pathway is dead, not that it does not help) and rgb.train.stat.feature_map_std must not collapse toward a constant map. A paired control -- this config with feature_conditioning.enabled=false -- is what makes the comparison attributable.