…evice/remote/strided

Flip TaskArgs Tensor→BufferRef end-to-end through the C++ dispatch; infer_deps keys a local ref by
canonical identity. Three-way consumer split (only the chip leaf rebuilds a C++ Tensor): chip →
materialize → run_from_blob; Python sub-worker → MappedArg (torch.frombuffer); nested L4→L3 → re-export
(no map on the forwarding hop, canonical identity invariant across the edge — frozen model §5/§8: an
L4-owned backing forwarded L4→L3→L2 keeps ONE identity at all three layers, so deps key on identity not
VA and aliases/retain-release do not split). Device-memory path (DEVICE_MALLOC materialize/wrap),
REMOTE_SIDECAR backend-rewrite at add_ref, and strided-ref → strided-Tensor materialization. Proven
L3→L2 + L4→L3→sub in sim and onboard a2a3.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… L2 in-process materialization

C-phase resource APIs (Worker.alloc_shared_tensor, Orchestrator.alloc_child_tensor); docs/
buffer-handle-abi.md (identity invariant across every edge; re-export preserves it); migrate the C++
GoogleTest orchestrator/scheduler/remote/child_memory tests to the BufferRef ABI; remove TensorTaskArgs
(fold onto ChipStorageTaskArgs); restore device-pointer provenance over the BufferRef wire; relax
create_buffer to L2 and materialize BufferRef args in-process on an L2 run (uniform user API at every
level).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Worker owns and manages the device-memory lifecycle (malloc / free / copy_to / copy_from /
alloc_child_tensor, keyed by owner_worker_id); the Orchestrator's memory APIs become thin forwards. The
raw-pointer orch.malloc / orch.device_handle are removed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…alars; orch.alloc→handle

Complete the Tensor removal from the user-facing path:
- ChipStorageTaskArgs is chip-only (ChipWorker._run_slot); Worker.run at L2 takes a BufferRef TaskArgs
  or None (legacy passthrough removed); migrate the remaining wire-TaskArgs pyut + remote
  materialization + test_callable_identity + scene-rehost UT fakes off add_tensor.
- Delete the create_host_buffer / MAP_HOST subsystem (superseded by create_buffer + lazy ref).
- handle.ref(dtype=) accepts a DataType enum; Python sub callables receive scalars (MappedArgs).
- orch.alloc returns a BufferHandle over a single identity-keyed Orchestrator::alloc; finish Tensor
  removal from the L3-L2 message queue / region.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Two O(1) warm-path static checks per the frozen model (worker-memory-model §1.4/§6, bufferhandle-abi
§3/§4.1), no ABI/wire change:

- Capability matrix: BufferHandleDescriptor.__post_init__ rejects an unsupported
  address_space×backend_kind (the four illegal combos HOST×VMM_WINDOW / HOST×DEVICE_MALLOC /
  DEVICE×FORK_SHM / DEVICE×POSIX_SHM). Runs on both owner construction and wire decode, so a bad combo
  fails before dispatch and cannot ride the wire. REMOTE_SIDECAR rows stay allowed here (its P1 blanket
  reject is ImportRegistry.materialize's job).
- access⊆granted: TaskArgs.add_ref (C++ binding) rejects an arg whose TensorArgType needs access the
  handle does not grant (INPUT→READ, OUTPUT_EXISTING→WRITE, INOUT→READWRITE). Catches e.g. a READ-only
  FORK_SHM COW buffer tagged OUTPUT_EXISTING (a forked child's writes would silently not reach the
  parent).

Not implemented here (already covered / P2, per the task doc): device→owner-chip topology
(_child_prov_check_dispatch), overlapping-alias (infer_deps serializes by identity), unknown
abi/backend/REMOTE_SIDECAR reject (already), stale-generation/ABA/visibility (P2).

Tests: test_descriptor_rejects_bad_capability_combo / accepts_legal_combos;
TestAddRefAccessSubset. Verified a2a3sim async_notify golden 0 (guards don't misfire on legit flows).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>