feat(disagg/mooncake): build_per_layer_context assembly (lever A, A3-step2)

Add MooncakeKVManager.build_per_layer_context: assembles a PerLayerTransferContext
from the SAME CP-filtered (prefill_kv_indices, dst_kv_indices) the post-forward
transfer uses — so the bytes moved are byte-identical to the monolithic path, and
the CP owner mapping is NOT re-derived (eliminating the #1 correctness risk). It
mirrors the MLA branch of _send_kvcache_generic exactly (get_mla_kv_ptrs_with_pp +
group_concurrent_contiguous + build_layer_blocks, verified set_transfer_blocks-
identical). Returns None for MHA / unregistered decode / empty owned set.

Unit-tested (4 cases): per-layer address correctness + the None guards. Remaining
A3-step3: call this in the send/scheduler flow (register before forward, finish
after, skip the main-KV monolithic send), then output-equality + TTFT verification.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/srt/disaggregation/mooncake/conn.py

@@ -349,6 +349,50 @@ class MooncakeKVManager(CommonKVManager):
             batch_ids.append(batch_id)
         return self.engine.wait_batch_transfers(batch_ids)
 
+    def build_per_layer_context(
+        self,
+        mooncake_session_id: str,
+        prefill_kv_indices: npt.NDArray[np.int32],
+        dst_kv_indices: npt.NDArray[np.int32],
+    ):
+        """Build a PerLayerTransferContext (lever A) for one request's main-KV pages,
+        from the SAME CP-filtered (prefill_kv_indices, dst_kv_indices) the post-forward
+        transfer would use — so the bytes moved are identical to the monolithic path.
+        The caller supplies the indices (reusing the send() CP filter, so this does NOT
+        re-derive the CP owner mapping — the #1 correctness risk). Mirrors the MLA
+        branch of _send_kvcache_generic exactly. Returns None if not applicable
+        (MHA / no decode registration yet / empty owned set)."""
+        if not self.is_mla_backend:
+            return None  # first lever-A impl targets MLA (the production GLM/NSA path)
+        reg = self.decode_kv_args_table.get(mooncake_session_id)
+        if reg is None or len(prefill_kv_indices) == 0:
+            return None
+        from sglang.srt.disaggregation.cp_per_layer_transfer import (
+            PerLayerTransferContext,
+            build_layer_blocks,
+        )
+
+        prefill_kv_blocks, dst_kv_blocks = group_concurrent_contiguous(
+            prefill_kv_indices, dst_kv_indices
+        )
+        src_kv_ptrs, dst_kv_ptrs, n_layers = self.get_mla_kv_ptrs_with_pp(
+            self.kv_args.kv_data_ptrs, reg.dst_kv_ptrs
+        )
+        item_lens = self.kv_args.kv_item_lens
+
+        def get_blocks(layer_id):
+            if layer_id >= n_layers:
+                return None
+            return build_layer_blocks(
+                src_kv_ptrs[layer_id],
+                dst_kv_ptrs[layer_id],
+                item_lens[layer_id],
+                prefill_kv_blocks,
+                dst_kv_blocks,
+            )
+
+        return PerLayerTransferContext(self.engine, mooncake_session_id, get_blocks)
+
     def _send_kvcache_generic(
         self,
         mooncake_session_id: str,