Anima loader: fix cosmetic tensor-count log + assert DiT fully materialized

from_single_file's converter pops the input state dict empty, so logging len(dit_sd)
after the load printed '0 tensors'. Capture counts before the load, and assert no
transformer param is left on the meta device (a half-mapped checkpoint would
otherwise crash later in pipe.to('cuda')).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -316,7 +316,8 @@ def _load_anima(entry):
         k[len(ADAPTER_PREFIX):]: v for k, v in raw.items()
         if k.startswith(ADAPTER_PREFIX)
     }
-    if not dit_sd:
+    n_dit, n_adapter = len(dit_sd), len(adapter_sd)  # capture now: from_single_file
+    if not dit_sd:                                    # pops dit_sd empty during convert
         raise RuntimeError(
             f"[anima] {entry['label']}: no '{DIT_PREFIX}*' keys in checkpoint "
             f"(got prefixes {sorted({k.split('.')[0] for k in raw})}); "
@@ -328,6 +329,14 @@ def _load_anima(entry):
         dit_sd, config=ANIMA_BASE, subfolder="transformer",
         torch_dtype=torch.bfloat16,
     )
+    # A half-mapped checkpoint would leave params on the meta device (later crashing
+    # in pipe.to("cuda")); assert the swap fully materialized so a key-format drift
+    # fails here with a clear message instead of silently degrading.
+    meta = [n for n, p in transformer.named_parameters() if p.device.type == "meta"]
+    if meta:
+        raise RuntimeError(
+            f"[anima] {entry['label']}: {len(meta)} transformer params unloaded "
+            f"(meta) after from_single_file, e.g. {meta[:3]} — key mapping is wrong.")
     pipe.update_components(transformer=transformer)
 
     # Finetuned LLM adapter -> the `text_conditioner` component. Its keys match the
@@ -345,8 +354,8 @@ def _load_anima(entry):
         print(f"[anima] {entry['label']}: no text_conditioner / adapter weights "
               f"(adapter_tensors={len(adapter_sd)}); using base adapter")
 
-    print(f"[anima] {entry['label']}: loaded finetuned DiT ({len(dit_sd)} tensors) "
-          f"+ adapter ({len(adapter_sd)} tensors)")
+    print(f"[anima] {entry['label']}: loaded finetuned DiT ({n_dit} tensors) "
+          f"+ adapter ({n_adapter} tensors)")
     pipe.to("cuda")
     return pipe