Add MPI direct L3 transport, runtime, and vector-add example.

CMakeLists.txt

@@ -67,6 +67,8 @@ add_custom_target(build_package_sim
 if(SKBUILD_MODE)
     install(DIRECTORY ${CMAKE_SOURCE_DIR}/src/
             DESTINATION simpler_setup/_assets/src)
+    install(DIRECTORY ${CMAKE_SOURCE_DIR}/cmake/
+            DESTINATION simpler_setup/_assets/cmake)
     install(DIRECTORY ${CMAKE_SOURCE_DIR}/build/lib/
             DESTINATION simpler_setup/_assets/build/lib
             OPTIONAL

docs/python-packaging.md

@@ -31,6 +31,7 @@ simpler_setup/            ← test framework + build/runtime assembly (simpler_s
     paged_attention.py    attention reference (used by multiple paged_attention tests)
   _assets/                (wheel-only, populated by CMake install)
     src/                  source tree mirror
+    cmake/                shared modules used by runtime compilation
     build/lib/            pre-built per-arch/platform/runtime .so/.o
 
 _task_interface.*.so      nanobind extension at site-packages root
@@ -69,10 +70,10 @@ NPU hardware (`a2a3`/`a5` with CANN toolkit).
 
 `simpler_setup.environment.PROJECT_ROOT` auto-detects between:
 
-- **Wheel install**: `simpler_setup/_assets/` exists → `PROJECT_ROOT = .../site-packages/simpler_setup/_assets`. The wheel's bundled `_assets/src/` and `_assets/build/lib/` provide everything needed at runtime.
+- **Wheel install**: `simpler_setup/_assets/` exists → `PROJECT_ROOT = .../site-packages/simpler_setup/_assets`. The wheel's bundled `_assets/src/`, `_assets/cmake/`, and `_assets/build/lib/` provide everything needed at runtime.
 - **Source tree / editable install**: `_assets/` doesn't exist → `PROJECT_ROOT = repo root`. Live `src/` and `build/lib/` are used.
 
-Anything that needs to find `src/`, `build/lib/`, or `build/cache/` MUST go through `simpler_setup.environment.PROJECT_ROOT` — never `Path(__file__).parent.parent...`.
+Anything that needs to find `src/`, `cmake/`, `build/lib/`, or `build/cache/` MUST go through `simpler_setup.environment.PROJECT_ROOT` — never `Path(__file__).parent.parent...`.
 
 ## Import rules
 

examples/workers/README.md

@@ -38,6 +38,7 @@ workers/
     global_tload_mixed_l3/  # Global CommDomain build + cross-machine peer TLOAD on both ranks
     compute_then_tload_mixed_l3/  # compute round on both L2s, then peer TLOAD through the same domain
     global_tload_mpirun_l3/ # one mpirun launches an L3 rank per machine; MPI descriptor exchange
+    vector_add_mpi_direct_l3/ # L4 joins MPI rank 0 and controls both real L3 ranks through direct P2P
 ```
 
 Why no `tensormap_and_ringbuffer/` layer? Because every example here hard-codes
@@ -67,6 +68,11 @@ parent owns a single `mpirun` that launches an L3 rank on each machine
 example's README for its extra prerequisites (`mpirun` + `mpi4py` on both
 machines).
 
+`vector_add_mpi_direct_l3` is the direct-MPI alternative: one supervisor
+launches L4 as MPI rank 0 and the two real L3 workers as ranks 1 and 2. It also
+needs `mpirun` and `mpi4py` on both machines, but task/control traffic is P2P
+between L4 and each L3 rather than passing through the PR2 group mailbox.
+
 ### What a new L4 example needs
 
 ```text

examples/workers/l4/vector_add_mpi_direct_l3/README.md

@@ -0,0 +1,68 @@
+# L4 direct MPI L3 vector add
+
+This example launches one static MPI world across two machines: rank 0 owns
+L4, rank 1 owns a real L3 on the L4 machine, and rank 2 owns a real L3 on the
+peer. L4 sends SLR3 task and control frames directly to both L3 ranks, and
+each L3 dispatches to its local L2 workers.
+
+```text
+machine A                                      machine B
+
+MPI rank 0 / L4 -------- direct MPI --------> MPI rank 2 / real L3
+       |
+       +------------- direct MPI -----------> MPI rank 1 / real L3
+```
+
+The example validates direct MPI startup and teardown, remote memory copy,
+L3-to-L2 vector-add execution, result checking, and Global CommDomain
+lifecycle. The vector-add kernel does not perform a peer `TLOAD`.
+
+## Prerequisites beyond the sibling example
+
+- `mpirun`/`mpiexec` and `mpi4py` must be installed on **both** machines and
+  built against the **same** MPI implementation.
+- `mpirun` executes one identical command line on both machines, so `--python`
+  must name an interpreter path valid on BOTH. Point it at a per-machine
+  launcher script installed at one shared absolute path; each machine's copy
+  sources CANN, enters that machine's checkout, and execs its `.venv` Python:
+
+```bash
+cat > /tmp/simpler-mpi-python <<'EOF'
+#!/usr/bin/env bash
+source /usr/local/Ascend/cann/set_env.sh
+cd /path/to/this/machines/simpler
+exec /path/to/this/machines/simpler/.venv/bin/python "$@"
+EOF
+chmod +x /tmp/simpler-mpi-python
+```
+
+## Run on the L4 parent
+
+`192.0.2.10` / `192.0.2.20` are documentation placeholders. Hosts must be
+numeric IPs; the local host is where ranks 0 and 1 run:
+
+```bash
+source .venv/bin/activate
+python -m examples.workers.l4.vector_add_mpi_direct_l3.main \
+  --local-host 192.0.2.10 --remote-host 192.0.2.20 \
+  --python /tmp/simpler-mpi-python \
+  --local-devices 0,1 --remote-devices 0,1 \
+  --mpirun-path "$(command -v mpiexec)" \
+  --launcher-family mpich
+```
+
+Success requires the process to return status 0 after printing:
+
+```text
+vector_add_mpi_direct_l3 passed
+```
+
+Any failed vector validation exits non-zero.
+
+## Running it in CI
+
+The network1 job's `network1-stage` action writes the per-machine launcher on both
+machines at one shared path and exports it as `NETWORK1_MPI_PYTHON`; the
+`test_vector_add_mpi_direct_l3.py` wrapper reads it together with `NETWORK1_LOCAL_IP`
+and the standard network1 fixtures, and skips when `mpirun`, `mpi4py`, or either
+variable is absent.

examples/workers/l4/vector_add_mpi_direct_l3/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) PyPTO Contributors.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# -----------------------------------------------------------------------------------------------------------
+"""L4-to-L3 direct MPI vector-add example."""