ADR-0026: DPPolicy intra-device only + ShardSpec structural coords

DPPolicy no longer carries a cross-SIP axis. SIP-level placement is
solely controlled by torch.ahbm.set_device(rank) (ADR-0024); DPPolicy
itself describes only the cube × PE layout within one SIP. ShardSpec
switches to structural (sip, cube, pe) coordinates; the flat pe_index
field/property is fully removed — silent drift between global-flat and
SIP-local interpretations was a foot-gun flagged by ADR-0024 D11.

Breaking API (explicit TypeError / AttributeError):
- DPPolicy(sip=...) / DPPolicy(num_sips=...) -> TypeError
- ShardSpec.pe_index -> AttributeError
- ShardSpec(pe_index=...) -> TypeError
- resolve_dp_policy now takes target_sip= (required), no num_sips.

Downstream migration:
- PE allocator dict keyed by (sip, cube, pe) tuples, in both
  _ensure_allocators and _free_tensor. deploy_tensor uses tuple lookup.
- _create_tensor passes target_sip=current_sip; post-hoc pe_index
  shifting removed entirely.
- launch._compute_local_shape drops the dp.sip branch.
- Internal resolvers (column_wise / row_wise / replicate / tiled_*)
  return _LocalPeShard (cube-local identifier) instead of ShardSpec —
  resolve_dp_policy lifts them to full structural coords.

Tests:
- New tests/test_adr0026_dppolicy_intra_device.py (12 tests) pins the
  contract end-to-end.
- test_sip_parallel.py rewritten: SIP composition now modeled as two
  resolve_dp_policy(target_sip=...) calls (ADR-0024 launcher style).
- Call-site migration: test_tensor, test_va_integration, test_va_offset,
  test_runtime_api_tensor, test_tl_recv_async, test_ccl_* and benches
  gemm_single_pe, gpt3_qkv, va_offset_verify, ccl_allreduce (legacy
  branch) all use intra-device DPPolicy and structural ShardSpec.

Result: 523 passed, 1 strict xfail (ring_default_ws — unchanged
ADR-0024 Phase B blocker; architectural fix deferred to ADR-0027).

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

tests/test_va_offset.py

@@ -20,7 +20,7 @@ from kernbench.policy.address.allocator import AddressConfig, PEMemAllocator
 from kernbench.policy.address.pe_mmu import PeMMU
 from kernbench.policy.address.phyaddr import PhysAddr
 from kernbench.policy.address.va_allocator import VirtualAllocator
-from kernbench.policy.placement.dp import DPPolicy, column_wise
+from kernbench.policy.placement.dp import DPPolicy, resolve_dp_policy
 from kernbench.runtime_api.tensor import deploy_tensor
 from kernbench.sim_engine.engine import GraphEngine
 from kernbench.runtime_api.context import RuntimeContext
@@ -70,7 +70,7 @@ def _make_standalone(shape, num_pe=NUM_PE):
         sram_bytes_per_cube=32 * _MB,
     )
     allocators = {
-        i: PEMemAllocator(rack_id=0, sip_id=0, cube_id=0, pe_id=i, cfg=cfg)
+        (0, 0, i): PEMemAllocator(rack_id=0, sip_id=0, cube_id=0, pe_id=i, cfg=cfg)
         for i in range(num_pe)
     }
     va_alloc = VirtualAllocator(va_base=0x1_0000_0000, va_size=64 * _GB, page_size=4096)
@@ -110,7 +110,11 @@ def test_2d_va_translates_to_local_hbm():
     cols_per_pe = K // NUM_PE
     block_bytes = M * cols_per_pe * ELEM_BYTES
 
-    placement = column_wise(shape=(M, K), itemsize=ELEM_BYTES, num_pe=NUM_PE)
+    placement = resolve_dp_policy(
+        DPPolicy(cube="replicate", pe="column_wise"),
+        shape=(M, K), itemsize=ELEM_BYTES,
+        num_pe=NUM_PE, num_cubes=1, target_sip=0,
+    )
     handle = deploy_tensor(
         name="src", shape=(M, K), dtype="fp16",
         placement=placement, allocators=allocators, va_allocator=va_alloc,
@@ -178,7 +182,11 @@ def test_1d_va_translates_to_local_hbm():
     elems_per_pe = N_1D // NUM_PE
     block_bytes = elems_per_pe * ELEM_BYTES
 
-    placement = column_wise(shape=(1, N_1D), itemsize=ELEM_BYTES, num_pe=NUM_PE)
+    placement = resolve_dp_policy(
+        DPPolicy(cube="replicate", pe="column_wise"),
+        shape=(1, N_1D), itemsize=ELEM_BYTES,
+        num_pe=NUM_PE, num_cubes=1, target_sip=0,
+    )
     handle = deploy_tensor(
         name="src_1d", shape=(N_1D,), dtype="fp16",
         placement=placement, allocators=allocators, va_allocator=va_alloc,
@@ -207,7 +215,9 @@ def test_1d_e2e_completes():
         correlation_id="vo6", spec=graph.spec,
     )
 
-    dp = DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")
+    # ADR-0026: DPPolicy is intra-device only; SIP scoping comes from the
+    # RuntimeContext's target_device. This 1D e2e runs on a single SIP.
+    dp = DPPolicy(cube="column_wise", pe="column_wise")
     src = ctx.zeros((N_1D,), dtype=DTYPE, dp=dp, name="src_1d")
     dst = ctx.empty((N_1D,), dtype=DTYPE, dp=dp, name="dst_1d")