Add fresh 24-GPU run checkpoint step 3500

Validated BF16 RGB+velocity weights and SHA-256 provenance. No deletions.

wan2.1_sim_real_24gpu_fresh_20260921/step-3500.json

@@ -0,0 +1,40 @@
+{
+  "checkpoint_contents": "DiT weights only; optimizer/RNG/training state not included",
+  "dtype": "bfloat16",
+  "export_recovery": {
+    "control_all_slices_bitwise_hash_match": true,
+    "control_rank_slices": 31272,
+    "control_step": 4200,
+    "control_tensors": 1303,
+    "reconstruction": "FP32 optimizer masters rounded to BF16, ZeRO padding trimmed"
+  },
+  "filename": "step-3500.safetensors",
+  "initialization": "Original pretrained Wan base; fresh optimizer at step 0; new velocity channels zero-initialized",
+  "model": "Wan2.1-I2V-14B-480P with RGB+velocity DiT expansion",
+  "note": "Separate fresh run, not a continuation of the previously published step-5000 checkpoint.",
+  "output_latent_channels": 32,
+  "parameters": 16395534208,
+  "patch_embedding_input_channels": 52,
+  "run": "fresh-base-24h100-ga1-sim5-real8-vel05-seed42-20260921",
+  "sha256": "57fe2f7735d5bc27dde7824d45525650c49fecd98ccb7d84e188ed181e4bc929",
+  "size_bytes": 32791205048,
+  "step": 3500,
+  "tensors": 1303,
+  "training": {
+    "data_seed": 42,
+    "dataset_repeat": 5,
+    "devices": 24,
+    "global_batch_size": 24,
+    "gradient_accumulation_steps": 1,
+    "height": 480,
+    "learning_rate": 1e-05,
+    "max_steps": 5000,
+    "microbatch_per_gpu": 1,
+    "num_frames": 81,
+    "real_dataset_repeat": 8,
+    "resume_from": null,
+    "resume_training_state": null,
+    "vel_loss_weight": 0.5,
+    "width": 832
+  }
+}

wan2.1_sim_real_24gpu_fresh_20260921/step-3500.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57fe2f7735d5bc27dde7824d45525650c49fecd98ccb7d84e188ed181e4bc929
+size 32791205048