Train: MultiSet + Muses Val V2 + Sky-SDF 16ch + C-RADIOv4 Alignment (REPA) 63b68ae2

Copy of "Train: MultiSet + Muses Val V2 + Sky-SDF 16ch + C-RADIOv4" (config 121, v2) with C-RADIOv4 moved from input to teacher: feature conditioning OFF, train-only REPA-style feature alignment ON. Data, losses, LRs and validation are otherwise unchanged. WHAT IS NEW. model.feature_alignment pulls the shared encoder's mid-block output (tap A), through a train-only 3-layer MLP head (9.44 M parameters), toward the frozen C-RADIOv4-H features of the paired CLEAR image, spatially normalised as in iREPA (gamma 0.7), by per-token cosine at every timestep (uniform weight). The clear image is privileged information that exists only in training, so this gives the shared encoder a dense, label-free target that is defined over the sky and the far range too. Nothing ships at inference: the heads are off the forward path and the backbone is never loaded there. Conditioning stays off so that a result can be attributed to the alignment (plan §6.2, §8.3). See docs/feature_alignment.md and docs/plans/foundation_feature_alignment.md. DELTAS FROM CONFIG 121. (1) feature_conditioning_enabled=false. The block is kept so the later {conditioning} x {alignment} arms are one flip away; both blocks read feature_backbone and feature_autocast_dtype, as the model requires. (2) model.feature_alignment.* added: enabled=true, target=clear, taps=[mid], head=mlp, head_hidden_channels=2048, target_norm_gamma=0.7, timestep_weight=uniform, probe_only=false. (3) loss.lambda_feature_alignment=0.5. (4) training.feature_alignment_learning_rate=1e-4, feature_alignment_probe_steps=10000, feature_alignment_ramp_steps=4000, feature_alignment_stop_step=null. (5) run_name, job_name and output_dir carry an -align suffix. run.sh changes only in comments: its torch.hub cache check still applies, because train_joint.py loads the alignment backbone at startup. WHY 10000/4000 AND NOT THE REPO'S 2000/2000. The global LR warm-up scales the heads' group too. warmup_ratio 0.1 of 200 epochs x ~1322 steps/epoch (measured on the C-RADIOv4 run) is ~26.4k steps, starting at 0.1% of every LR, so a 2000-step head-only phase would give the heads ~4% of the LR it assumes. 10000 steps gives them about the budget of 2000 steps at 1e-4, and the ramp adds ~1800 more. The trunk feels the full alignment from step 14000 (~epoch 10.6, ~19 h at the ~5 s/step the conditioning run logs). ARMS (plan §8.3): this config with one parameter changed. Baseline: feature_alignment_probe_only=true (the trunk trains exactly as without alignment and the same cosines are logged). Align: the defaults. Control: feature_alignment_target=hazy. Keep the seed and step budget identical across arms. UNMEASURED. No GPU has run this pathway. lambda=0.5 is REPA's value and a calibration start, not a tuned weight. Read rgb.train.stat.feature_alignment_grad_share from step 14000, or during the ramp divide it by sys.train.feature_alignment_trunk_grad_scale. Set lambda so the full-strength share is 0.1-0.3, then fix it for the A/B. Peak memory and step time are unmeasured; the conditioning run, with the same single backbone forward, logs 79.6-83.3 GB used of 85.9 GB at batch 24. gpu_type still reads rtx_4090:1, so set the 80 GB GPU at launch, as for config 121. REQUIRES the alignment code in Euler's checkout. repa-align is merged into main at 54bdc11, so pull main there before launching. WATCH. rgb.train.stat.feature_alignment_cosine must rise during the head-only phase (flat means the heads are not learning) and should plateau before step 10000; if it does not, lengthen feature_alignment_probe_steps. The readouts that test the premise are rgb.train.stat.feature_alignment_cosine_t3 (t >= 810) and the validation cosine at the deployed first step, rgb.train.diag.feature_alignment_cosine in the val split. feature_alignment_grad_share appears only from step 10000.