Merge branch 'codex/canonical-complete-20260715'

Author: apstenku123

.github/workflows/ci-self-hosted.yml

@@ -26,8 +26,11 @@ jobs:
         shell: bash
         run: |
           set -euo pipefail
-          python_bin=/Volumes/external/sources/cppmega.mlx/.venv/bin/python
+          python_bin=/Volumes/external/sources/.venvs/cppmega.mlx/bin/python
           test -x "${python_bin}"
+          env -u PYTHONPATH -u PYTHONHOME -u VIRTUAL_ENV \
+            "${python_bin}" scripts/check_mlx_abi.py --json
+          uv pip check --python "${python_bin}"
           "${python_bin}" scripts/run_self_hosted_ci.py lane \
             --lane macos-mlx \
             --python "${python_bin}" \

.github/workflows/cuda-lane.yml

@@ -14,16 +14,60 @@ jobs:
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
 
-      - name: Set up Python
-        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-        with:
-          python-version: "3.13"
-
-      - name: Install dependencies
+      - name: Verify runner-provided CUDA environment
         run: |
-          python -m pip install --upgrade pip
-          pip install -e ".[dev]"
+          set -euo pipefail
+          : "${CPPMEGA_CUDA_PYTHON:?self-hosted CUDA runner must export CPPMEGA_CUDA_PYTHON}"
+          test -x "$CPPMEGA_CUDA_PYTHON"
+          cuda_info="$(env -u PYTHONPATH -u PYTHONHOME -u VIRTUAL_ENV \
+            "$CPPMEGA_CUDA_PYTHON" -c \
+            'import os,sys; print(os.path.realpath(sys.prefix)); print(os.path.realpath(sys.base_prefix))')"
+          cuda_prefix="${cuda_info%%$'\n'*}"
+          cuda_base_prefix="${cuda_info#*$'\n'}"
+          test -n "$cuda_prefix"
+          test -n "$cuda_base_prefix"
+          test "$cuda_prefix" != "$cuda_base_prefix"
+          test -f "$cuda_prefix/pyvenv.cfg"
+          python_abs="$(python3 -c 'import os,sys; print(os.path.abspath(sys.argv[1]))' "$CPPMEGA_CUDA_PYTHON")"
+          case "$python_abs" in
+            "$cuda_prefix/bin/"*) ;;
+            *) echo "selected CUDA Python is outside its venv prefix" >&2; exit 2 ;;
+          esac
+          workspace_real="$(cd "$GITHUB_WORKSPACE" && pwd -P)"
+          case "$cuda_prefix" in
+            "$workspace_real"|"$workspace_real"/*)
+              echo "refusing CUDA environment inside GITHUB_WORKSPACE" >&2
+              exit 2
+              ;;
+          esac
+          if git -C "$cuda_prefix" rev-parse --show-toplevel >/dev/null 2>&1; then
+            echo "refusing CUDA environment inside a Git checkout" >&2
+            exit 2
+          fi
+          uv lock --check --no-sources --project .
+          UV_PROJECT_ENVIRONMENT="$cuda_prefix" uv sync --project . --locked \
+            --no-sources --python "$CPPMEGA_CUDA_PYTHON" --group dev \
+            --extra path-c
+          "$CPPMEGA_CUDA_PYTHON" scripts/check_cuda_environment.py --json \
+            | tee "$RUNNER_TEMP/cppmega-cuda-environment.json"
+          uv pip check --python "$CPPMEGA_CUDA_PYTHON"
+
+      - name: Upload CUDA environment receipt
+        if: always()
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: cppmega-mlx-cuda-environment-${{ github.run_id }}-${{ github.run_attempt }}
+          path: ${{ runner.temp }}/cppmega-cuda-environment.json
+          if-no-files-found: warn
+          retention-days: 14
 
       - name: Run CUDA tests
         run: |
-          python -m pytest -m "cuda_hardware or cuda_required" -v --tb=short
+          set -euo pipefail
+          : "${CPPMEGA_CUDA_PYTHON:?self-hosted CUDA runner must export CPPMEGA_CUDA_PYTHON}"
+          selected="$RUNNER_TEMP/cppmega-cuda-selected.txt"
+          CPPMEGA_RUN_CUDA_TESTS=1 "$CPPMEGA_CUDA_PYTHON" -m pytest \
+            --collect-only -q -m "cuda_hardware or cuda_required" > "$selected"
+          grep -q '::' "$selected"
+          CPPMEGA_RUN_CUDA_TESTS=1 "$CPPMEGA_CUDA_PYTHON" -m pytest \
+            -m "cuda_hardware or cuda_required" -v --tb=short

.github/workflows/e2e-matrix.yml

@@ -16,8 +16,11 @@ on:
     branches: [main]
     paths:
       - "cppmega_v4/**"
+      - "cppmega_mlx/**"
       - "vbgui/**"
       - "tests/fixtures/build_e2e_matrix.py"
+      - "tests/fixtures/parquet/**"
+      - "tests/fixtures/tokenizers/**"
       - "scripts/install_self_hosted_e2e_python.sh"
       - ".github/workflows/e2e-matrix.yml"
   schedule:
@@ -54,7 +57,7 @@ jobs:
           node-version: "22"
       - name: Create isolated MLX Python
         run: |
-          base_python=/Volumes/external/sources/cppmega.mlx/.venv/bin/python
+          base_python=/Volumes/external/sources/.venvs/cppmega.mlx/bin/python
           test -x "$base_python"
           job_venv="$(mktemp -d "$RUNNER_TEMP/cppmega-mlx-e2e-${GITHUB_RUN_ID}-${GITHUB_JOB}.XXXXXX")"
           "$base_python" -m venv "$job_venv"
@@ -120,7 +123,7 @@ jobs:
         with: { node-version: "22" }
       - name: Create isolated MLX Python
         run: |
-          base_python=/Volumes/external/sources/cppmega.mlx/.venv/bin/python
+          base_python=/Volumes/external/sources/.venvs/cppmega.mlx/bin/python
           test -x "$base_python"
           job_venv="$(mktemp -d "$RUNNER_TEMP/cppmega-mlx-e2e-${GITHUB_RUN_ID}-${GITHUB_JOB}.XXXXXX")"
           "$base_python" -m venv "$job_venv"
@@ -178,7 +181,7 @@ jobs:
         with: { node-version: "22" }
       - name: Create isolated MLX Python
         run: |
-          base_python=/Volumes/external/sources/cppmega.mlx/.venv/bin/python
+          base_python=/Volumes/external/sources/.venvs/cppmega.mlx/bin/python
           test -x "$base_python"
           job_venv="$(mktemp -d "$RUNNER_TEMP/cppmega-mlx-e2e-${GITHUB_RUN_ID}-${GITHUB_JOB}.XXXXXX")"
           "$base_python" -m venv "$job_venv"

.gitignore

@@ -34,7 +34,7 @@ docs/upstream/tilelang_metal_mamba3/captures/
 .claude/
 
 # Large real-data samples copied from GB10 for local-only smoke tests.
-data/parquet_samples/
+data/parquet_samples
 
 # Beads / Dolt files (added by bd init)
 .beads/

MANIFEST.in

@@ -0,0 +1,16 @@
+include README.md
+include pyproject.toml
+include setup.py
+include cppmega_mlx/tokenizer/tokenizer_contract_v1.json
+include cppmega_mlx/tokenizer/tokenizer.json
+include cppmega_mlx/data/domain_schema_v1.json
+include cppmega_v4/nn/_external/VENDORED_MANIFEST.json
+global-exclude *.so *.dylib *.metallib
+recursive-include cppmega_mlx/training/native_optim \
+    CMakeLists.txt \
+    bindings.cpp \
+    fused_8bit.cpp \
+    fused_8bit.h \
+    fused_8bit.metal
+recursive-include cppmega_v4/widget/static *.css *.mjs
+global-exclude *.so *.dylib *.metallib

README.md

@@ -222,7 +222,8 @@ matrix at
 2. dsa_splitk budget gate — tiled Q-cache for `AH≥8` production shapes
 3. ATEN_DISPATCH `_scaled_dot_product_flash_attention_for_cpu` wiring (FA TileLang kernel exists at `_kernels/flash_attention.py`)
 4. reduce_prod `vectorize_loop.cc` / `storage_rewrite.cc` mul-kind handling
-5. `scripts/check_mlx_abi.sh` to catch venv-vs-brew dylib mismatch (host venv `mlx.core.so` was built against older `libmlx.dylib` — silent test skips)
+5. `scripts/check_mlx_abi.sh` to catch `mlx`/`mlx-metal` drift before kernel
+   tests (run it against the isolated env; the checkout `.venv` is shared)
 
 NVFP4 on MLX research is committed at
 [`tilelang/docs/research/nvfp4_mlx_metal.md`](https://github.com/DatasunriseOU/tilelang/blob/main/docs/research/nvfp4_mlx_metal.md):