fix remote train processed cache path

scripts/remote_train.sh

@@ -41,8 +41,8 @@ export DEEPGENOPIX_NUM_WORKERS=4
 
 MODE=$([ "${RESUME:-0}" = "1" ] && echo resume || echo train)
 
-# Canonical cache key — emitted by `deepgenopix matrix --json` for each preset.
-PROCESSED_SIGNATURE="$(uv run python -c "
+# Local processed directory signature used by the trainer.
+LOCAL_PROCESSED_SIGNATURE="$(uv run python -c "
 from deepgenopix.notebook_support import build_experiment_config
 cfg = build_experiment_config('$PRESET')
 parts = [
@@ -53,12 +53,15 @@ parts = [
     f'tf{str(cfg.test_frac).replace(\".\", \"p\")}',
     f'seed{cfg.split_seed}',
     f'mf{cfg.min_family_size}',
-    'ds$DATA_REVISION_SHORT',
 ]
 print('_'.join(parts))
 ")"
+
+# HF dataset cache key. Include the dataset revision so corrected split uploads
+# cannot hit an LMDB produced from a stale dataset snapshot.
+PROCESSED_SIGNATURE="${LOCAL_PROCESSED_SIGNATURE}_ds${DATA_REVISION_SHORT}"
 PIXEL_STRIDE="$(uv run python -c "from deepgenopix.notebook_support import build_experiment_config; print(build_experiment_config('$PRESET').pixel_stride_bp)")"
-PROCESSED_DIR="data/processed/te_visuals/auto"
+PROCESSED_DIR="data/processed/te_visuals/$LOCAL_PROCESSED_SIGNATURE"
 mkdir -p "$PROCESSED_DIR"
 
 # Try to pull an existing ETL cache (LMDB + registry + classes) by signature.