[feat] delta dump supports zch

tzrec/main.py

@@ -809,7 +809,6 @@ def train_and_evaluate(
             train_config.delta_embedding_dump_config,
             pipeline_config.model_dir,
             device,
-            pipeline_config.feature_configs,
         )
 
     dense_optim_cls, dense_optim_kwargs = optimizer_builder.create_dense_optimizer(

tzrec/protos/train.proto

@@ -79,7 +79,10 @@ message FeatureStoreConfig {
 }
 
 message DeltaEmbeddingDumpConfig {
-    // MC/ZCH features are not supported; use dynamicemb for delta dump.
+    // ZCH features publish raw feature ids: touched and newly admitted ids
+    // with their trained rows, and evicted ids with the fallback row the
+    // model now serves for them (FeatureStore has no delete, so the old
+    // vector must be overwritten explicitly).
     // Dump touched ids and their latest embedding every N training steps. Do not
     // set this together with dump_interval_minutes.
     // Larger intervals retain a longer id window in memory; auto compaction reduces