Preserve draft KV across CP HiCache hits

Cache-hit prefill can skip draft forward for the prefix while PD transfer still reads draft KV for that same prefix.  CP HiCache therefore needs to persist draft/MTP KV alongside target KV instead of relying on whatever remains in the draft GPU pool.

Constraint: CP HiCache is host-only here; storage backends remain unsupported for CP shared KV.

Constraint: CP shared KV must keep owner-page semantics and avoid falling back to full KV on every rank.

Rejected: Recompute cached-prefix draft KV during prefill | loses the HiCache benefit and reintroduces the large hidden/KV footprint.

Rejected: Change PD transfer to skip draft prefix KV | decode still needs draft cache continuity for MTP acceptance.

Confidence: medium

Scope-risk: moderate

Directive: Keep target and draft CP HiCache metadata/load/write/evict paths in lockstep; changing one without the other can silently reduce MTP accept length.

Tested: Remote g0034 container /sgl-workspace/sglang-tai: python3 -m pytest -q test/registered/unit/managers/test_hicache_controller_cp.py test/registered/unit/mem_cache/test_cp_hicache_metadata.py => 58 passed, 3 warnings

Not-tested: Full multi-node HiCache+MTP serving benchmark and accept-length recovery.

python/sglang/srt/disaggregation/prefill.py

@@ -853,9 +853,18 @@ class SchedulerDisaggregationPrefillMixin:
             )
             return
         prefill_queue = getattr(self, "disagg_prefill_bootstrap_queue", None)
+        has_draft_pool = (
+            getattr(prefill_queue, "draft_token_to_kv_pool", None) is not None
+        )
+        prefix_len = len(getattr(req, "prefix_indices", ()))
+        host_hit_length = int(getattr(req, "host_hit_length", 0) or 0)
+        draft_prefix_overlap = max(0, min(end_idx, prefix_len) - start_idx)
         _cp_draft_shared_kv_debug(
             "prefill_send_kv_chunk rid=%s room=%s start_idx=%s end_idx=%s "
-            "last_chunk=%s page_size=%s pages=%s state_pages=%s has_draft_pool=%s",
+            "last_chunk=%s page_size=%s pages=%s state_pages=%s "
+            "has_draft_pool=%s prefix_len=%s host_hit_length=%s "
+            "cache_protected_len=%s extend_input_len=%s fill_len=%s "
+            "origin_input_len=%s already_computed=%s draft_prefix_overlap=%s",
             req.rid,
             req.bootstrap_room,
             start_idx,
@@ -864,6 +873,27 @@ class SchedulerDisaggregationPrefillMixin:
             page_size,
             _seq_summary(page_indices),
             _seq_summary(state_indices),
-            getattr(prefill_queue, "draft_token_to_kv_pool", None) is not None,
+            has_draft_pool,
+            prefix_len,
+            host_hit_length,
+            getattr(req, "cache_protected_len", None),
+            getattr(req, "extend_input_len", None),
+            len(req.fill_ids),
+            len(req.origin_input_ids),
+            getattr(req, "already_computed", None),
+            draft_prefix_overlap,
         )
+        if has_draft_pool and draft_prefix_overlap > 0:
+            _cp_draft_shared_kv_debug(
+                "prefill_send_cachehit_draft_prefix rid=%s room=%s "
+                "draft_prefix_overlap=%s prefix_len=%s host_hit_length=%s "
+                "start_idx=%s end_idx=%s note=transfer_reads_draft_pool_for_cached_prefix",
+                req.rid,
+                req.bootstrap_room,
+                draft_prefix_overlap,
+                prefix_len,
+                host_hit_length,
+                start_idx,
+                end_idx,
+            )
         req.disagg_kv_sender.send(page_indices, state_indices)