Kernel-launch sync (ADR-0009 D5) and IPCQ drain at inbound (ADR-0023)

- KernelLaunchMsg gains target_start_ns: IO_CPU stamps a global barrier
  (max path latency across every target PE), M_CPU passes it through,
  PE_CPU yields until it before recording pe_exec_start. Every PE in a
  launch begins kernel execution at the same env.now regardless of its
  dispatch path length — eliminates per-PE dispatch-offset artifact in
  cross-PE and cross-cube latency measurements.

- PE_DMA._handle_ipcq_inbound now pays Transaction.drain_ns at the top,
  matching the terminal-drain behavior of ComponentBase._forward_txn for
  every non-IPCQ Transaction. SRC-side tl.send stays fire-and-forget
  (sender doesn't yield on sub_done); tl.recv now blocks until bytes
  have actually drained into its inbox.

- ComponentContext: new compute_path_latency_ns helper + node_overhead_ns
  field populated by GraphEngine.

- tests/test_kernel_launch_sync.py: asserts all PEs in one launch
  produce identical pe_exec_ns for a no-op kernel (zero spread).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/kernbench/components/builtin/pe_dma.py

@@ -186,13 +186,37 @@ class PeDmaComponent(PeEngineBase):
     # ── IPCQ inbound (fabric → PE_DMA → MemoryStore + PE_IPCQ) ──────
 
     def _handle_ipcq_inbound(self, env: simpy.Environment, txn: Any) -> Generator:
-        """At destination PE_DMA: atomically write data and forward metadata.
+        """At destination PE_DMA: pay terminal drain, then atomically write
+        data and forward metadata.
+
+        ADR-0023 D9 (drain at inbound terminal): the Transaction carries
+        ``drain_ns = nbytes / bottleneck_bw_on_path`` stamped by the sender
+        PE_DMA. Like every other Transaction terminal in the simulator (see
+        ``ComponentBase._forward_txn``), this drain must be paid when the
+        Transaction reaches its destination. SRC-side ``tl.send`` is
+        fire-and-forget — it never yields on ``sub_done`` — so paying the
+        drain here does NOT delay the sender. What it DOES delay is the
+        IpcqMetaArrival forwarded below: that delay is the only signal
+        ``tl.recv`` on DST blocks on, which is exactly the desired
+        semantics — "send dispatches and returns; recv waits until the
+        bytes have actually landed in its inbox".
+
+        The drain MUST be paid before the atomic block — inserting a yield
+        inside would break invariant I6.
 
         I6 (MUST): no SimPy yield between MemoryStore.write and the
         IpcqMetaArrival put into PE_IPCQ.
         """
         from kernbench.common.ipcq_types import IpcqMetaArrival
 
+        # Pay terminal BW drain before the atomic write/metadata forward.
+        # Without this, IPCQ effectively got fabric bandwidth for free at
+        # the terminal (only intermediate-hop overhead_ns was charged),
+        # making IPCQ lower than raw DMA at large sizes in benchmarks.
+        drain = getattr(txn, "drain_ns", 0.0)
+        if drain > 0:
+            yield env.timeout(drain)
+
         token = txn.request
 
         # ── ATOMIC: do not introduce yield between these two operations ──