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.

astrai/inference/cache/strategy.py

@@ -89,6 +89,19 @@ class Allocator:
             if idx in self._lru:
                 self._lru.move_to_end(idx)
 
+    def clear_cached(self) -> int:
+        """Release every unreferenced LRU page back to the free pool."""
+        with self._lock:
+            cached = list(self._lru)
+            self._lru.clear()
+            for idx in cached:
+                if self._refs[idx] != 0:
+                    raise RuntimeError("Cannot invalidate a referenced cache page")
+                if self.on_evict:
+                    self.on_evict(idx)
+                self._free_mask |= 1 << idx
+            return len(cached)
+
 
 class RadixNode:
     """A page-aligned edge in the CPU-side prefix radix trie."""
@@ -200,6 +213,10 @@ class AllocationStrategy(ABC):
         start: int,
     ) -> None: ...
 
+    def invalidate_cache(self) -> int:
+        """Drop reusable KV entries after an inference weight update."""
+        return 0
+
 
 class ContiguousStrategy(AllocationStrategy):
     """Static contiguous allocation: slots are pre-assigned at pool init.
@@ -316,3 +333,8 @@ class PagedStrategy(AllocationStrategy):
         full = len(prompt_ids) // self._page_size
         for i in range(start, min(full, len(state.pages))):
             self._prefix.record(state.pages[i], prompt_ids, i)
+
+    def invalidate_cache(self) -> int:
+        if self._prefix is None:
+            return 0
+        return self._alloc.clear_cached()