fix: resolve audited training, import, and serving bugs

- shard the Muon Newton-Schulz orthogonalization over the FSDP mesh instead of partial local slices
- import HF checkpoints faithfully: per-head RoPE permutation for q/k projections and qk-norm, qwen3, shared experts, and qk-norm before RoPE (changes numerics for existing use_qk_norm checkpoints)
- make preprocessing and resume self-contained: backfill realigned bucket keys by semantics (masks ones, rest zeros) and snapshot tokenizer files into every checkpoint
- keep RL consistent: sync the offline GRPO old_model each optimizer step and validate online strategies through a public one-off-rollout hook that leaves the replay cache untouched
- fix streaming serving: withhold partial tool-call prefixes with a stream-end flush, stream tool-call arguments from the raw source span, and terminate SSE frames with a blank line
- fix sampling semantics: capture logprobs before top-k/top-p mutate logits in place and detect greedy pipelines polymorphically instead of isinstance bookkeeping

astrai/trainer/rollout.py

@@ -551,3 +551,16 @@ class RolloutRunner:
         # cache publication. Reward scoring itself intentionally remains
         # outside the policy lock because it may call an external service.
         return self.generator.with_policy_snapshot(commit)
+
+    def evaluate(self, batch: Dict) -> RolloutResult:
+        """One-off rollout + scoring that leaves the replay cache untouched.
+
+        Used by validation on online strategies: the training cache, its
+        cadence counter, and the cache key stay intact, so evaluation
+        prompts never disturb the rollout replay schedule.
+        """
+        raw = self.generator.generate(batch)
+        self._validate_policy_version(raw)
+        scored = self._score(raw)
+        self._validate_policy_version(scored)
+        return scored