feat(sc): recover replay buffer from native TQ checkpoints

examples/configs/grpo_math_1B.yaml

@@ -496,6 +496,10 @@ data_plane:
   storage_capacity: 1000000          # max samples retained per partition
   num_storage_units: 2               # storage shards
   claim_meta_poll_interval_s: 0.5    # blocking-claim poll cadence
+  # SingleController only: save native TQ state. Required when trainer
+  # checkpointing is enabled with a replay-checkpoint-capable sampler;
+  # supported samplers restore from metadata-only replay indexes.
+  checkpointing_enabled: false
   global_segment_size: 549755813888  # 512 GiB — used when backend == "mooncake_cpu"
   local_buffer_size:   68719476736   # 64 GiB  — used when backend == "mooncake_cpu"
   # observability:                    # NotRequired

examples/configs/grpo_math_1B_megatron_single_controller.yaml

@@ -21,7 +21,6 @@ async_rl:
   # Enable per-rollout diagnostic prints (prompt content / completion previews).
   diagnostics: false
 
-# SC does not support checkpointing yet.
 checkpointing:
   enabled: false
   checkpoint_dir: results/grpo-single-controller
@@ -55,6 +54,8 @@ logger:
 
 data_plane:
   enabled: true
+  # Required shadow snapshot: a save failure aborts checkpoint finalization.
+  checkpointing_enabled: true
 
 cluster:
   gpus_per_node: 2