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_tensor.py

@@ -2,11 +2,13 @@ import pytest
 
 from kernbench.policy.address.allocator import AddressConfig, AllocationError, PEMemAllocator
 from kernbench.policy.placement.dp import (
+    DPPolicy,
     ShardSpec,
     column_wise,
-    tiled_column_major,
     replicate,
+    resolve_dp_policy,
     row_wise,
+    tiled_column_major,
     tiled_row_major,
 )
 from kernbench.runtime_api.kernel import (
@@ -40,9 +42,9 @@ _CFG = AddressConfig(
 )
 
 
-def _make_allocators(num_pe: int = 8) -> dict[int, PEMemAllocator]:
+def _make_allocators(num_pe: int = 8) -> dict[tuple[int, int, int], PEMemAllocator]:
     return {
-        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)
     }
 
@@ -133,7 +135,7 @@ def test_column_wise_placement():
     assert len(shards) == 8
     expected_nbytes = 1024 * 64 * 2  # 128 KB
     for i, s in enumerate(shards):
-        assert s.pe_index == i
+        assert s.local_pe == i
         assert s.nbytes == expected_nbytes
     # offsets are contiguous
     assert shards[0].offset_bytes == 0
@@ -151,7 +153,7 @@ def test_row_wise_placement():
     assert len(shards) == 8
     expected_nbytes = 128 * 512 * 2  # 128 KB
     for i, s in enumerate(shards):
-        assert s.pe_index == i
+        assert s.local_pe == i
         assert s.nbytes == expected_nbytes
     assert shards[0].offset_bytes == 0
     assert sum(s.nbytes for s in shards) == 1024 * 512 * 2
@@ -166,7 +168,7 @@ def test_replicate_placement():
     assert len(shards) == 8
     full_nbytes = 1024 * 512 * 2  # 1 MB
     for i, s in enumerate(shards):
-        assert s.pe_index == i
+        assert s.local_pe == i
         assert s.nbytes == full_nbytes
         assert s.offset_bytes == 0  # each is a full copy
 
@@ -188,10 +190,10 @@ def test_tiled_column_major():
     # tile (m=0,k=0) → PE0, tile (m=0,k=1) → PE1, ..., (m=0,k=3) → PE3
     # tile (m=1,k=0) → PE4, tile (m=1,k=1) → PE5, ..., (m=1,k=3) → PE7
     # tile (m=2,k=0) → PE0, ...
-    assert shards[0].pe_index == 0
-    assert shards[1].pe_index == 1
-    assert shards[7].pe_index == 7
-    assert shards[8].pe_index == 0  # wraps around
+    assert shards[0].local_pe == 0
+    assert shards[1].local_pe == 1
+    assert shards[7].local_pe == 7
+    assert shards[8].local_pe == 0  # wraps around
     # total coverage
     assert sum(s.nbytes for s in shards) == 1024 * 512 * 2
 
@@ -212,10 +214,10 @@ def test_tiled_row_major():
     # tile (m=0,k=0) → PE0, tile (m=1,k=0) → PE1, ..., (m=3,k=0) → PE3
     # tile (m=0,k=1) → PE4, tile (m=1,k=1) → PE5, ..., (m=3,k=1) → PE7
     # tile (m=0,k=2) → PE0, ...
-    assert shards[0].pe_index == 0
-    assert shards[1].pe_index == 1
-    assert shards[7].pe_index == 7
-    assert shards[8].pe_index == 0  # wraps around
+    assert shards[0].local_pe == 0
+    assert shards[1].local_pe == 1
+    assert shards[7].local_pe == 7
+    assert shards[8].local_pe == 0  # wraps around
     # total coverage
     assert sum(s.nbytes for s in shards) == 1024 * 512 * 2
 
@@ -226,7 +228,11 @@ def test_tiled_row_major():
 def test_deploy_tensor_hbm():
     """Deploy with column_wise placement → TensorHandle with valid PA shards."""
     allocs = _make_allocators()
-    placement = column_wise(shape=(1024, 512), itemsize=2, num_pe=8)
+    placement = resolve_dp_policy(
+        DPPolicy(cube="replicate", pe="column_wise"),
+        shape=(1024, 512), itemsize=2,
+        num_pe=8, num_cubes=1, target_sip=0,
+    )
     th = deploy_tensor(
         name="W",
         shape=(1024, 512),
@@ -253,7 +259,7 @@ def test_deploy_tensor_hbm():
 def test_deploy_tensor_tcm():
     """Deploy with TCM → uses pe_tcm_addr allocation."""
     allocs = _make_allocators()
-    placement = [ShardSpec(pe_index=0, offset_bytes=0, nbytes=256)]
+    placement = [ShardSpec(sip=0, cube=0, pe=0, offset_bytes=0, nbytes=256)]
     th = deploy_tensor(
         name="small",
         shape=(128,),
@@ -271,7 +277,7 @@ def test_deploy_tensor_overflow():
     """Allocation exceeding PE HBM capacity raises AllocationError."""
     allocs = _make_allocators()
     # 6 GB per PE slice, try to allocate 7 GB
-    big_shard = ShardSpec(pe_index=0, offset_bytes=0, nbytes=7 * _GB)
+    big_shard = ShardSpec(sip=0, cube=0, pe=0, offset_bytes=0, nbytes=7 * _GB)
     with pytest.raises(AllocationError):
         deploy_tensor(
             name="toobig",