Latency model: HBM PC striping + chunk-loop drain (ADR-0033)

Previous model double-counted slow-upstream paths (e.g., 64KB via UCIe
128 GB/s was ~2x pessimistic). HBM CTRL now distributes bursts across
8 pseudo-channels via global round-robin, with per-chunk commit timing
that pipelines correctly against the bottleneck link's data arrival.

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

src/kernbench/sim_engine/transaction.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+from collections.abc import Iterator
 from dataclasses import dataclass, field
 from typing import Any
 
@@ -47,3 +48,46 @@ class Transaction:
             is_response=self.is_response,
             result_data=self.result_data,
         )
+
+    def into_flits(self, flit_bytes: int) -> Iterator[Flit]:
+        """Decompose this Transaction's payload into Flits (ADR-0033 D1).
+
+        Yields one Flit per ``flit_bytes`` of payload. The final flit may
+        carry fewer bytes when ``nbytes`` is not a multiple of ``flit_bytes``;
+        that flit has ``is_last=True``. Transactions with ``nbytes <= 0``
+        yield no flits.
+
+        All yielded Flits share a reference to this Transaction.
+        """
+        if self.nbytes <= 0 or flit_bytes <= 0:
+            return
+        n_full = self.nbytes // flit_bytes
+        remainder = self.nbytes % flit_bytes
+        n_total = n_full + (1 if remainder else 0)
+        for i in range(n_total):
+            size = flit_bytes if i < n_full else remainder
+            yield Flit(
+                txn=self,
+                flit_index=i,
+                flit_nbytes=size,
+                is_last=(i == n_total - 1),
+            )
+
+
+@dataclass
+class Flit:
+    """Atomic wire transport unit (ADR-0033 D1).
+
+    Carries a slice of a parent Transaction's payload. The wire
+    (``engine._wire``) decomposes Transactions into Flits on first
+    transport; downstream wires pass Flits through with their own
+    ``bw_gbs`` delay.
+
+    Phase 2 constraint: ``flit_bytes`` MUST be a multiple of HBM
+    ``burst_bytes`` (default they are equal). See ADR-0033 D1.
+    """
+
+    txn: Transaction       # parent transaction reference
+    flit_index: int        # 0..n_flits-1
+    flit_nbytes: int       # bytes carried (usually flit_bytes; last may be smaller)
+    is_last: bool          # True for the terminating flit