Replace DMD EMA with aligned step-480 adapter

Publish checkpoint-1200 EMA from remind-dmd15k-sf4s-a1-b64s4-0919, reproduced with the locked W&B UI-step 480 validation protocol on the TF+2000 base.

README.md

@@ -99,14 +99,16 @@ python inference.py \
 
 The public inference recipe uses 81 frames at 832×480 and 16 fps, seven
 three-latent-frame chunks, and the shifted four-step schedule
-`[1000, 937, 833, 625]`. Camera examples use the bundled pair-fixed GT camera
+`[1000, 938, 833, 625]`. Camera examples use the bundled pair-fixed GT camera
 trajectories.
 
 ## Checkpoint lineage and validation
 
 - ReMind-5B generator: pair-fixed 480p checkpoint after an additional 2,000
   teacher-forcing iterations; this is the frozen base used by the current DMD line.
-- DMD EMA adapter: last complete SF-DMD checkpoint 2500.
+- DMD EMA adapter: checkpoint 1200 (W&B UI step 480) from
+  `remind-dmd15k-sf4s-a1-b64s4-0919`, validated against the locked eight-case
+  raw/EMA suite with the TF+2000 base.
 - The ReMind-5B checkpoint loads with 100% key coverage.
 
 Integrity hashes are provided in `SHA256SUMS`.

ReMind-5b-dmd-ema.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8e786351c0375f5fe0b5c44659272e4aa55ef395fae79860cfab42786150d598
+oid sha256:35aac81065da0e4139897405ecd2aa5403ff703eb753f80a89b895ef5a5dab59
 size 1478584376

SHA256SUMS

@@ -1,2 +1,2 @@
-8e786351c0375f5fe0b5c44659272e4aa55ef395fae79860cfab42786150d598  ReMind-5b-dmd-ema.safetensors
+35aac81065da0e4139897405ecd2aa5403ff703eb753f80a89b895ef5a5dab59  ReMind-5b-dmd-ema.safetensors
 fda02ffe06f98d97ca5c42af8b0eef1d627ebc01d95c5364020e7ad8261a4683  ReMind-5B.safetensors