Harden CPU AI image builds

CHANGELOG.md

@@ -7,6 +7,12 @@
 
 # Changelog
 
+## Sprint 172 CPU AI image build hardening (2026-07-12)
+
+- Reordered the Unraid all-in-one Docker build so backend source changes reuse the Python/GIS/AI dependency layer.
+- Pinned the opt-in CPU runtime to PyTorch 2.13.0 and torchvision 0.28.0 from the official CPU wheel index, avoiding unused CUDA runtime packages while preserving the currently validated framework versions.
+- Kept AI dependencies opt-in and model weights local-only; no API, migration, model activation or inference contract changed.
+
 ## Sprint 171 Positive AOI expansion and split safety (2026-07-12)
 
 - Added four explicit, real-reference Kempen training AOIs: Olen, Lille, Oud-Turnhout and Kasterlee center.

backend/tests/test_docker_runtime_config.py

@@ -265,6 +265,21 @@ def test_backend_docker_context_contains_gis_import_smoke_script() -> None:
     assert (ROOT / "backend" / "scripts" / "gis_import_smoke.py").exists()
 
 
+def test_all_in_one_dockerfile_caches_dependencies_and_uses_cpu_torch_for_ai_runtime() -> None:
+    dockerfile = (ROOT / "deploy" / "unraid" / "Dockerfile.all-in-one").read_text(encoding="utf-8")
+
+    metadata_copy_index = dockerfile.index("COPY backend/pyproject.toml backend/README.md /app/")
+    dependency_install_index = dockerfile.index('RUN /usr/bin/python3.11 -m venv /opt/geointel/venv \\')
+    backend_copy_index = dockerfile.index("COPY backend/ /app/")
+    smoke_index = dockerfile.index("RUN python scripts/gis_import_smoke.py")
+
+    assert metadata_copy_index < dependency_install_index < backend_copy_index < smoke_index
+    assert "GEOINTEL_TORCH_INDEX_URL=https://download.pytorch.org/whl/cpu" in dockerfile
+    assert "GEOINTEL_TORCH_VERSION=2.13.0" in dockerfile
+    assert "GEOINTEL_TORCHVISION_VERSION=0.28.0" in dockerfile
+    assert '--index-url "$GEOINTEL_TORCH_INDEX_URL"' in dockerfile
+
+
 def test_unraid_deploy_passes_ai_build_arg_and_yolo_runtime_env() -> None:
     deploy_ps1 = (ROOT / "scripts" / "deploy_tower.ps1").read_text(encoding="utf-8")
     deploy_sh = (ROOT / "scripts" / "deploy_tower.sh").read_text(encoding="utf-8")

deploy/unraid/Dockerfile.all-in-one

@@ -9,6 +9,9 @@ RUN npm run build
 FROM postgres:16-bookworm AS runtime
 
 ARG GEOINTEL_INSTALL_AI=false
+ARG GEOINTEL_TORCH_INDEX_URL=https://download.pytorch.org/whl/cpu
+ARG GEOINTEL_TORCH_VERSION=2.13.0
+ARG GEOINTEL_TORCHVISION_VERSION=0.28.0
 
 ENV GEOINTEL_ENV=production \
     GEOINTEL_API_PREFIX=/api/v1 \
@@ -44,15 +47,25 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 
 WORKDIR /app
 
-COPY backend/ /app/
-COPY fixtures/ /app/fixtures/
+COPY backend/pyproject.toml backend/README.md /app/
+COPY backend/app/__init__.py /app/app/__init__.py
 
 RUN /usr/bin/python3.11 -m venv /opt/geointel/venv \
     && pip install --no-cache-dir --upgrade pip setuptools \
+    && if [ "$GEOINTEL_INSTALL_AI" = "true" ]; then \
+         pip install --no-cache-dir \
+           --index-url "$GEOINTEL_TORCH_INDEX_URL" \
+           "torch==$GEOINTEL_TORCH_VERSION" \
+           "torchvision==$GEOINTEL_TORCHVISION_VERSION"; \
+       fi \
     && extras=".[gis]" \
     && if [ "$GEOINTEL_INSTALL_AI" = "true" ]; then extras=".[gis,ai]"; fi \
-    && pip install --no-cache-dir "$extras" \
-    && python scripts/gis_import_smoke.py \
+    && pip install --no-cache-dir "$extras"
+
+COPY backend/ /app/
+COPY fixtures/ /app/fixtures/
+
+RUN python scripts/gis_import_smoke.py \
     && python scripts/yolo_preflight.py --json >/tmp/geointel-yolo-preflight.json \
     && rm -f /etc/nginx/sites-enabled/default \
     && mkdir -p /app/storage /run/nginx /var/log/nginx

deploy/unraid/README.md

@@ -96,6 +96,11 @@ GIS-only image. Set `GEOINTEL_INSTALL_AI=true`, mount models through
 when you have a local model file.
 The AI-enabled image installs PyTorch/Ultralytics plus the native OpenCV runtime
 libraries needed for Ultralytics imports; it still never downloads model weights.
+The documented CPU runtime installs pinned `torch==2.13.0` and
+`torchvision==0.28.0` from `https://download.pytorch.org/whl/cpu`, avoiding the
+unused CUDA runtime wheels included by the general Linux package index. The
+Dockerfile copies dependency metadata before backend source, so normal code-only
+redeploys can reuse the expensive dependency layer.
 `YOLO_CONFIG_DIR` defaults to `/app/storage/ultralytics`, a writable persistent
 path, so Ultralytics settings do not fall back to root user config directories.
 

docs/CODEX_EXECUTION_LOG.md

@@ -7015,3 +7015,21 @@ Open:
 ## Next recommended pass
 
 - Finish the inactive candidate, run fixed-threshold positive evidence and split pure-empty/sparse-context background matrices, and preserve the current production default unless the promotion report passes every gate.
+# Sprint 172 - CPU AI image build hardening
+
+## What changed
+
+- Reordered `deploy/unraid/Dockerfile.all-in-one` so `pyproject.toml` and minimal package metadata are installed before the complete backend source is copied.
+- Code-only backend changes can now reuse the expensive GIS/AI dependency layer; dependency metadata changes still invalidate it.
+- The opt-in CPU AI build now installs the same validated PyTorch `2.13.0` / torchvision `0.28.0` versions from the official CPU wheel index before installing the `ai` extra.
+- Kept the full GIS and YOLO import/preflight smoke after the complete backend source copy.
+
+## Validation so far
+
+- RED: the new Docker ordering/CPU-wheel regression test failed against the old Dockerfile.
+- GREEN: `python -m pytest backend/tests/test_docker_runtime_config.py -q`: 26 passed.
+- `python -m pip index versions` confirmed `torch 2.13.0+cpu` and `torchvision 0.28.0+cpu` are available from the configured CPU index for the workstation platform.
+
+## Remaining validation
+
+- Build the AI-enabled all-in-one image on Tower after the current inactive model training run finishes, verify Torch reports a CPU build and rerun live migration/browser smokes before replacing the runtime.

docs/DEPENDENCY_POLICY.md

@@ -43,6 +43,18 @@ AI dependencies remain separate in the `ai` optional dependency group and must
 not be installed by the default Docker backend image unless an explicit AI image
 or profile is introduced later.
 
+The opt-in Unraid all-in-one AI build is CPU-oriented because its documented
+runtime sets `YOLO_DEVICE=cpu`. It installs the pinned PyTorch/torchvision pair
+from PyTorch's CPU wheel index before installing the `ai` extra. This avoids
+shipping unused CUDA runtime libraries. The index and versions remain explicit
+Docker build arguments so a future, separately validated GPU image can override
+them without changing the base dependency group.
+
+Docker dependency metadata is copied before application source. Backend source
+changes therefore reuse the dependency layer while changes to `pyproject.toml`
+still invalidate it correctly. Runtime GIS and YOLO import/preflight smokes run
+after the complete backend source is copied.
+
 ## Approved AI
 
 - PyTorch