PD-sep server-side profiling: vLLM patches + per-request breakdown

Instrumentation patches (microbench/patches/):
  - pd_profile.py: shared event emitter (VLLM_PD_PROFILE_LOG env var)
  - apply_patches.py: idempotent patch installer for mooncake_connector.py
    and scheduler.py, marks insertions with # PD_PROFILE_PATCH
  - analyze_events.py: joins per-process JSONL event logs by transfer_id
    into per-request phase durations

Seven events captured per request:
  D_get_num_matched → P_zmq_received → P_prefill_done →
  P_rdma_start → P_rdma_end → D_recv_complete → D_request_promoted

Driver fix (microbench/lifecycle/driver.py):
  seed_prefix_cache now sends via the proxy URL so P and D both cache
  the seeded prefix with matching block hashes. Previously seeding D
  directly produced different block hashes than the proxy-routed
  measurement requests, making incremental transfer impossible.

Real breakdown (fig_breakdown_real.png, server_breakdown.csv, n=93):
  prefill_compute  620 ms median (95% of overhead)
  rdma_transfer     42 ms median (~71 Gbps effective)
  other overhead    10 ms median (dispatch + params + signal + promote)

Mooncake transfer is NOT the bottleneck. Even with bulk RDMA the
transfer cost is <10% of prefill cost for Qwen3-30B-A3B on H20.

microbench/lifecycle/driver.py

@@ -88,7 +88,14 @@ def make_new_tokens_prompt(num_tokens: int, unique_id: str) -> str:
 async def seed_prefix_cache(
     client: httpx.AsyncClient, url: str, model: str, num_tokens: int, session_id: str
 ) -> bool:
-    """Send a request to D to warm its prefix cache with num_tokens of context.
+    """Warm BOTH P and D prefix caches by sending the seed through the PD-sep proxy.
+
+    Sending directly to D would only warm D's cache but produce block hashes that
+    don't match what P later produces (different tokenization path). Sending through
+    the proxy makes P do the prefill (P caches), pushes KV to D (D caches with
+    matching hashes), so subsequent requests with the same prefix get incremental
+    transfer on both sides.
+
     Returns True if successful.
     """
     if num_tokens == 0:
@@ -318,11 +325,11 @@ async def main():
             print(f"Prior context C={C} tokens")
             print(f"{'='*60}")
 
-            # Seed D's prefix cache
+            # Seed BOTH P and D prefix caches via the proxy
             if C > 0:
-                print(f"  Seeding D prefix cache with {C} tokens...")
+                print(f"  Seeding P+D prefix caches with {C} tokens via proxy...")
                 success = await seed_prefix_cache(
-                    client, seed_endpoint, args.model, C, args.session_id
+                    client, pdsep_url, args.model, C, args.session_id
                 )
                 if not success:
                     print(f"  SKIP all configs with C={C} (cache seed failed)")