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

tests/trainer/test_rollout.py

@@ -388,6 +388,21 @@ def test_rollout_runner_cache_returns_stale_flag(device):
     assert fresh2 is False
 
 
+def test_rollout_runner_evaluate_leaves_cache_untouched(device):
+    runner, _ = _make_runner(device, rollout_interval=10)
+    batch = _make_instruction_batch()
+    cached, _ = runner(batch)
+
+    eval_batch = _make_instruction_batch(n=1)
+    result = runner.evaluate(eval_batch)
+
+    assert result.rewards.shape == result.responses.shape[:2]
+    replayed, fresh = runner(batch)
+    assert replayed is cached
+    assert fresh is False
+    assert runner._steps_since_rollout == 0
+
+
 def test_rollout_runner_tags_generation_version_and_preserves_cached_behavior(device):
     runner, _ = _make_runner(device, rollout_interval=100)
     batch = _make_instruction_batch(n=1)