Add --verify-data CLI flag, Tensor.data property, parallel DataExecutor

- CLI: --verify-data flag enables Phase 2 data verification (ADR-0020)
- Tensor.data: returns actual numpy values (verify-data) or zeros placeholder
- Tensor.__repr__: shows value summary or data=N/A (placeholder)
- DataExecutor: ThreadPoolExecutor for same-timestamp parallel op execution
- BenchResult.engine: exposes op_log/memory_store for Phase 2 access

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/kernbench/runtime_api/bench_runner.py

@@ -62,6 +62,7 @@ def run_bench(
             correlation_id=correlation_id,
             trace=None,
             traces=collected_traces,
+            engine=engine,
         )
 
     if completion_policy == CompletionPolicy.LAST_SUBMITTED:
@@ -69,7 +70,7 @@ def run_bench(
         completion, trace = engine.get_completion(last)
         return BenchResult(
             completion=completion, correlation_id=correlation_id,
-            trace=trace, traces=collected_traces,
+            trace=trace, traces=collected_traces, engine=engine,
         )
 
     if completion_policy == CompletionPolicy.ALL_OK_FAIL_FAST:
@@ -80,11 +81,11 @@ def run_bench(
             if not c.ok:
                 return BenchResult(
                     completion=c, correlation_id=correlation_id,
-                    trace=last_trace, traces=collected_traces,
+                    trace=last_trace, traces=collected_traces, engine=engine,
                 )
         return BenchResult(
             completion=Completion(ok=True), correlation_id=correlation_id,
-            trace=last_trace, traces=collected_traces,
+            trace=last_trace, traces=collected_traces, engine=engine,
         )
 
     # LAST_COMPLETED placeholder (needs engine support for timing). Fall back.
@@ -92,5 +93,5 @@ def run_bench(
     completion, trace = engine.get_completion(last)
     return BenchResult(
         completion=completion, correlation_id=correlation_id,
-        trace=trace, traces=collected_traces,
+        trace=trace, traces=collected_traces, engine=engine,
     )

src/kernbench/runtime_api/context.py

@@ -314,6 +314,7 @@ class RuntimeContext:
         t._handle = handle
         import weakref
         t._ctx_ref = weakref.ref(self)
+        t._memory_store = getattr(self.engine, "_memory_store", None)
         self._tensors.append(weakref.ref(t))
 
         # Install VA→PA mappings via fabric MmuMapMsg

src/kernbench/runtime_api/tensor.py

@@ -5,6 +5,8 @@ import weakref
 from dataclasses import dataclass
 from typing import Literal
 
+import numpy as np
+
 from kernbench.policy.address.allocator import PEMemAllocator
 from kernbench.policy.placement.dp import DPPolicy, ShardSpec
 from kernbench.runtime_api.kernel import TensorArg, TensorArgShard
@@ -50,6 +52,20 @@ def dtype_itemsize(dtype: str) -> int:
     return _DTYPE_ITEMSIZE[dtype]
 
 
+_NUMPY_DTYPE = {
+    "f16": np.float16, "fp16": np.float16, "float16": np.float16,
+    "f32": np.float32, "fp32": np.float32, "float32": np.float32,
+    "bf16": np.float16,
+    "i8": np.int8, "int8": np.int8,
+    "i16": np.int16, "int16": np.int16,
+    "i32": np.int32, "int32": np.int32,
+}
+
+
+def _numpy_dtype(dtype: str) -> np.dtype:
+    return np.dtype(_NUMPY_DTYPE.get(dtype, np.float16))
+
+
 def deploy_tensor(
     *,
     name: str,
@@ -129,6 +145,7 @@ class Tensor:
         self._dp_metadata: DPMetadata | None = None
         self._handle: TensorHandle | None = None
         self._ctx_ref: weakref.ref | None = None  # set by RuntimeContext
+        self._memory_store = None  # set by RuntimeContext when enable_data=True
 
     def __del__(self) -> None:
         if self._ctx_ref is None or self._handle is None:
@@ -137,6 +154,28 @@ class Tensor:
         if ctx is not None:
             ctx._free_tensor(self)
 
+    def __repr__(self) -> str:
+        parts = [f"tensor(name={self.name}, shape={self.shape}, dtype={self.dtype}"]
+        if self._memory_store is not None and self._handle is not None:
+            arr = self.data
+            parts.append(f", mean={float(arr.mean()):.4g}, norm={float(np.linalg.norm(arr)):.4g}")
+        else:
+            parts.append(", data=N/A (placeholder)")
+        parts.append(")")
+        return "".join(parts)
+
+    @property
+    def data(self) -> np.ndarray:
+        """Tensor data as numpy array. Returns actual values when enable_data=True,
+        zeros placeholder otherwise (like an uninitialized tensor)."""
+        if self._memory_store is not None and self._handle is not None:
+            shard = self._handle.shards[0]
+            try:
+                return self._memory_store.read("hbm", shard.pa, shape=self.shape, dtype=self.dtype)
+            except KeyError:
+                pass
+        return np.zeros(self.shape, dtype=_numpy_dtype(self.dtype))
+
     @property
     def itemsize(self) -> int:
         return dtype_itemsize(self.dtype)

src/kernbench/runtime_api/types.py

@@ -12,6 +12,7 @@ class BenchResult:
     correlation_id: str
     trace: Trace | None = None
     traces: list[dict] | None = None
+    engine: object | None = None  # GraphEngine ref for Phase 2 data access
 
     def summary_text(self) -> str:
         if self.completion.ok: