fix(orchestrator): restore zero-advantage filtering flag

docs/algorithms.md

@@ -443,9 +443,9 @@ demo_key = "demonstration"
 
 Scoring runs before curriculum admission, so a rollout that is later rejected still costs its reference compute.
 
-The orchestrator filters samples with no training signal. This includes samples with zero advantage on all RL tokens. Samples that still carry CE or reference-KL components are retained. Filtering an RL token also removes its trainer/inference mismatch-KL contribution.
+By default, the orchestrator filters samples with no training signal. This includes samples with zero advantage on all RL tokens. Samples that still carry CE or reference-KL components are retained. Filtering an RL token also removes its trainer/inference mismatch-KL contribution. Set `orchestrator.train.filter_zero_advantages = false` to retain zero-advantage RL tokens and their mismatch-KL contribution.
 
-`orchestrator.constant_trainer_batch_size` defaults to `true`. The orchestrator filters samples before they count toward the batch target. It collects replacements, so rollout-based batches contain `orchestrator.batch_size` training traces. Set the option to `false` to filter after collection without replacement. This setting can improve orchestrator throughput, but it produces smaller trainer batches.
+`orchestrator.constant_trainer_batch_size` defaults to `true`. When zero-advantage filtering is enabled, the orchestrator filters samples before they count toward the batch target. It collects replacements, so rollout-based batches contain `orchestrator.batch_size` training traces. Set the option to `false` to filter after collection without replacement. This setting can improve orchestrator throughput, but it produces smaller trainer batches.
 
 ## Curricula
 

packages/prime-rl-configs/src/prime_rl/configs/orchestrator.py

@@ -308,6 +308,9 @@ class TrainConfig(BaseConfig):
     sampling: TrainSamplingConfig = TrainSamplingConfig()
     """Shared training sampling configuration."""
 
+    filter_zero_advantages: bool = True
+    """Remove zero-advantage RL tokens before shipping samples to the trainer."""
+
     @model_validator(mode="after")
     def resolve_env_defaults(self):
         """Resolve per-env overrides: inherit group-level sampling (the worker ``pool``

src/prime_rl/orchestrator/train_sink.py

@@ -293,7 +293,7 @@ async def process_group(self, group_id: str) -> None:
                         "it requires vLLM's native sampling-mask capture (>= 0.28)."
                     )
                 stamp_loss_routing(sample, env.algorithm.action_loss_type)
-            if self.config.constant_trainer_batch_size:
+            if self.config.train.filter_zero_advantages and self.config.constant_trainer_batch_size:
                 samples = [sample for sample in samples if _prune_zero_advantages(sample)]
             if samples:
                 samples_by_trace[trace.id] = samples
@@ -383,7 +383,7 @@ def process_batch(self) -> TrainBatch:
         for trace_id in selected_ids:
             del self.pending_batch[trace_id]
 
-        if not self.config.constant_trainer_batch_size:
+        if self.config.train.filter_zero_advantages and not self.config.constant_trainer_batch_size:
             selected_by_trace = {
                 trace_id: [sample for sample in samples if _prune_zero_advantages(sample)]
                 for trace_id, samples in selected