Add SLAC FEL model files and preliminary workflow

.gitignore

@@ -1,5 +1,12 @@
 # Data and configs
-*.toml
+#*.toml
+examples/slac_fel/data/
+# DS_Store files
+.DS_Store
+.idea
+
+double_check_data.ipynb
+uv.lock
 
 # Logging files
 *.db

CLAUDE.md

@@ -75,6 +75,16 @@ Note: `EnsembleDetector` is recognized by the loader but raises `NotImplementedE
 | `kfac_online` | KFAC approximation | kfac_lambda, kfac_ema_decay |
 | `none` | No-op (skip CL) | (none) |
 
+Replay of historical data is handled in `BaseUpdater.fwd_bwd()` and gated by
+`[continual_learning] mix_historic_data` (default `false`). When enabled and the
+harness supplies historical dataloaders, half of each stream is combined into a
+single forward/backward pass, so the sample count per step stays at
+`train.batch_size` whether or not mixing is on (an undersized historical batch is
+topped up from the current one). `base`, `ewc_online`, and `kfac_online` inherit this. `jvp_reg` ignores the flag: it
+overrides `fwd_bwd()` and mixes the two streams itself (it needs a current-only
+gradient as the JVP tangent direction), so it calls `super().fwd_bwd(batch)` without
+the historical batch. `none` skips training entirely.
+
 ### Coding Conventions
 - Python 3.13+, type hints everywhere
 - Formatting: ruff format, ruff check, mypy

examples/mnist/model.py

@@ -142,21 +142,29 @@ def get_hist_dataloaders(
         self,
     ) -> Tuple[Optional[DataLoader], Optional[DataLoader]]:
         """
-        If no history yet: add current drift to history and return (None, None).
-        Else: return loaders over ConcatDataset of prior drifts, then append current drift to history.
-        Effective dataset length = len(aug_history) * len(full_split).
+        If no history yet: return (None, None).
+        Else: return loaders over a replay buffer spanning EVERY prior cumulative
+        regime (0 … k-1), not just the immediately-previous one. Each past regime
+        `i` is the composition ``FixedAffine(aug_history[:i])``, matching the data
+        that was actually streamed at window i-1. Concatenating all of them keeps
+        the model anchored to the earliest regimes so CL adaptation does not
+        specialise to the most recent drift alone.
+        Effective dataset length = (len(aug_history) - 1) * len(full_split).
         """
         if self.task_counter == 1:
             return None, None
 
-        # Concatenate FULL train/val views for each historical drift
+        # One FULL train/val view per prior cumulative regime: regimes 0 … k-1.
+        # aug_history[-1] is the current regime, so we stop before it.
         train_views = [
             TransformedView(
-                self.ds_train, x_transform=FixedAffine(self.aug_history[:-1])
+                self.ds_train, x_transform=FixedAffine(self.aug_history[:i])
             )
+            for i in range(1, len(self.aug_history))
         ]
         val_views = [
-            TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:-1]))
+            TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:i]))
+            for i in range(1, len(self.aug_history))
         ]
 
         ds_hist_train: ConcatDataset[Any] = ConcatDataset(train_views)