feat: version rollout weight updates

Track a monotonic policy version across optimizer steps, scheduler updates, and rollout results. Serialize synchronous generation with weight acknowledgements and invalidate reusable prefix KV entries so cached samples remain attributable to the behavior policy that generated them.

tests/inference/test_cache.py

@@ -1,5 +1,6 @@
 """Unit tests for inference cache components."""
 
+import pytest
 import torch
 
 from astrai.inference.cache import (
@@ -435,6 +436,26 @@ def test_page_pool_prefix_hit_populates_request_mapping():
     )
 
 
+def test_task_cache_invalidation_drops_cross_version_prefix_hits():
+    pool = _make_paged_pool_ps64(page_size=2, max_seq_len=8, n_tokens=16)
+    task_cache = _make_task_cache(pool)
+    prompt = [11, 12, 13, 14]
+
+    assert task_cache.task_alloc("first", prompt)
+    task_cache.task_record_hashes("first", prompt)
+    task_cache.task_free("first")
+    assert task_cache.task_alloc("cached", prompt)
+    assert task_cache.task_cached("cached") == len(prompt)
+
+    with pytest.raises(RuntimeError, match="while tasks are active"):
+        task_cache.invalidate_cache()
+
+    task_cache.task_free("cached")
+    assert task_cache.invalidate_cache() == 2
+    assert task_cache.task_alloc("after_update", prompt)
+    assert task_cache.task_cached("after_update") == 0
+
+
 def test_page_pool_paged_ps64_bind_roundtrip():
     pool = _make_paged_pool_ps64(n_layers=1, n_kv_heads=2, head_dim=4)
     task_cache = _make_task_cache(pool)

tests/inference/test_scheduler.py

@@ -545,6 +545,34 @@ def test_run_batch_empty_prompts(device):
         scheduler.stop()
 
 
+def test_scheduler_weight_versions_are_monotonic_and_acknowledged(device):
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        assert scheduler.policy_version == 0
+        assert scheduler.update_weights(1) == 1
+        assert scheduler.policy_version == 1
+        assert scheduler.get_stats()["policy_version"] == 1
+        assert scheduler.update_weights(1) == 1
+        with pytest.raises(ValueError, match="cannot move backwards"):
+            scheduler.update_weights(0)
+        with pytest.raises(ValueError, match="non-negative integer"):
+            scheduler.update_weights(True)
+    finally:
+        scheduler.stop()
+
+
+def test_scheduler_rejects_weight_update_with_queued_tasks(device):
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    task_id = scheduler.add_task("queued")
+    try:
+        with pytest.raises(RuntimeError, match="while tasks are queued"):
+            scheduler.update_weights(1)
+        scheduler.remove_task(task_id)
+        assert scheduler.update_weights(1) == 1
+    finally:
+        scheduler.stop()
+
+
 def test_run_batch_too_long_prompt_skipped(device):
     """A prompt longer than max_seq_len yields an empty result slot."""
     scheduler, _tok, _model = _make_real_scheduler(device)