feat: replace non-overlapping windows with sliding-window sequence sampling

- Remove sample-level shuffle before transforms (broke SimulateEvents)
- Add _sliding_window_fn: yields overlapping sequences with configurable stride
- Add sequence-level shuffle after grouping (preserves temporal coherence)
- Add sliding_window_stride to TrainConfig (stride=1 for full overlap)
- Update create_train/val_loader and train.py to pass stride
- AGENTS.md: document known issues (cross-shard boundary, SimulateEvents state)
- AGENTS.md: add cuda:7 device preference

Generated by Mistral Vibe (deepseek-v4-flash).
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>

src/velocity_prediction/config.py

@@ -39,10 +39,11 @@ VAL_SCENES = [
     # "indoor_forward_3", "indoor_forward_9", "indoor_forward_10",  # Easy
 ]
 TEST_SCENES = [
+    "indoor_forward_9","indoor_forward_3",
     # "indoor_forward_7",      # Hard 室内
     # "outdoor_forward_1",     # Easy 室外
     # "outdoor_forward_5"      # Hard 室外
-    "indoor_forward_3", "indoor_forward_9", "indoor_forward_10",  # Easy
+    # "indoor_forward_3", "indoor_forward_9", "indoor_forward_10",  # Easy
 ]
 
 
@@ -93,7 +94,7 @@ class ModelConfig:
 class TrainConfig:
     seq_len: int = 8                    # frames per training sequence
     batch_size: int = 32
-    epochs: int = 100
+    epochs: int = 300
     lr: float = 1e-3
     weight_decay: float = 1e-5
     lr_scheduler_step: int = 30
@@ -101,6 +102,9 @@ class TrainConfig:
     num_workers: int = 4
     seed: int = 42
 
+    # Sliding window: stride=1 → full overlap, stride=seq_len → non-overlapping
+    sliding_window_stride: int = 1
+
     # Event simulation
     event_threshold: float = 0.1
     event_use_log: bool = True