fix(ai-bundles): lock the numpy-1.x ABI closure so the OCR bundle can't strand scipy (#437)

* fix(ai-bundles): lock the numpy-1.x ABI closure so the OCR bundle can't strand scipy

The OCR bundle installs paddleocr[doc-parser] 3.4, whose dependency closure drags
numpy 1.26.4 up to 2.5.1 and pulls scipy/scikit-learn/pandas wheels built against
the numpy 2.x ABI. build-bundle.sh re-pinned only numpy (basePackages), so those
numpy-2.x wheels stayed behind; the by-dir-name site-packages diff then shipped
them, and once merged onto the numpy==1.26.4 base they raise "numpy.dtype size
changed" on import.

Because the dispatcher pre-imports every ML library at startup and disables all AI
after 5 crashes in 60s, one stranded scipy takes down every AI tool, not just OCR
(observed on a CPU host: remove-background worked before the OCR bundle and broke
after). All-7 installs escaped it through last-writer-wins ordering; a subset
install did not, which is why it surfaced only intermittently.

Fix: add a manifest "constraints" list (numpy, scipy, scikit-learn, scikit-image,
pandas pinned to numpy-1.x-ABI versions) and apply it via PIP_CONSTRAINT to every
bundle pip install, so no bundle can pull a numpy-2.x wheel. paddleocr 3.4.1 still
resolves cleanly under the lock and the pinned stack imports without ABI error on
numpy 1.26.4 (validated on py3.12). Also import scipy/sklearn in the OCR path of
verify-bundle.sh so CI catches this class in isolation, and add a manifest
regression test.

Note: the published bundles must be rebuilt and republished (ai-bundles.yml) for
this to reach already-installed bases.

Claude-Session: https://claude.ai/code/session_01UvVCMNUBrgpghk8gye5gav

* chore(ai-bundles): sync OCR manifest sha256 to the rebuilt numpy-1.x bundles

Rebuilt the OCR bundle for both arches with the numpy-1.x-ABI constraints from
this PR and republished the tars to deepsafe/feature-bundles/v2.0.0, then updated
the baked manifest sha256 and sizes so installs verify against the fixed archives:

  amd64-gpu  5.93 GB  sha 2a00a3184f6a635f1fa9ae2a6517ad740a11f9e5ff58c098d2fd369a2bb1e16b
  arm64-cpu  1.98 GB  sha 6868c264069dcb74c6675c0b1f58dc1c9f60d9aa4459725e3dbde07a99a6a09a

Both tars ship scipy 1.12.0 / scikit-learn 1.4.2 / pandas 2.2.2 (numpy-1.x-ABI)
and zero numpy-2.x wheels, verified by listing the archive contents.

Stopgap note: these tars were built against the ghcr.io latest base (the 2.0.0
image is not published to GHCR), so they are not byte-identical to what the CI
build will produce. When ai-bundles.yml rebuilds at the 2.0.0 release, it will
mint fresh sha256 values and this manifest must be re-synced to them.

Claude-Session: https://claude.ai/code/session_01UvVCMNUBrgpghk8gye5gav

docker/build-bundle.sh

@@ -65,6 +65,32 @@ if '${BUNDLE_ID}' not in m['bundles']:
     sys.exit(1)
 "
 
+# ── Step 0: Write pip constraints (numpy-1.x-ABI closure lock) ────────────
+# Bundle installs below can otherwise pull the latest transitive scientific
+# stack: numpy 2.x plus scipy/scikit-learn/scikit-image/pandas wheels built
+# against the numpy 2.x ABI (e.g. paddleocr[doc-parser] drags numpy 1.26.4 ->
+# 2.5.1 + scipy 1.18). The Step 4 re-pin snaps numpy back to 1.26.4 but leaves
+# those numpy-2.x wheels behind; the Step 6 diff (by dir name, not version)
+# then ships them, and once installed they strand on the numpy==1.26.4 base and
+# raise "numpy.dtype size changed" on import, crashing the dispatcher for
+# EVERY AI tool (e.g. rembg's scipy), not just this bundle. Applying the
+# manifest's constraints to every pip install keeps the whole closure on
+# numpy-1.x-ABI versions so no bundle can strand a numpy-2.x wheel.
+CONSTRAINTS_FILE="/tmp/bundle-constraints.txt"
+python3 -c "
+import json
+with open('${MANIFEST}') as f:
+    m = json.load(f)
+with open('${CONSTRAINTS_FILE}', 'w') as f:
+    f.write('\n'.join(m.get('constraints', [])) + '\n')
+"
+if [[ -s "${CONSTRAINTS_FILE}" ]] && [[ -n "$(tr -d '[:space:]' < "${CONSTRAINTS_FILE}")" ]]; then
+  export PIP_CONSTRAINT="${CONSTRAINTS_FILE}"
+  echo "  Build constraints: $(tr '\n' ' ' < "${CONSTRAINTS_FILE}")"
+else
+  echo "  No build constraints in manifest"
+fi
+
 # Clean previous build artifacts
 rm -rf "${MODELS_DIR}" "${BUILD_DIR}"
 mkdir -p "${MODELS_DIR}" "${BUILD_DIR}/site-packages" "${BUILD_DIR}/models" "${OUTPUT_DIR}"

docker/feature-manifest.json

@@ -3,6 +3,13 @@
   "imageVersion": "2.0.0",
   "pythonVersion": { "amd64": "3.12", "arm64": "3.11" },
   "basePackages": ["numpy==1.26.4", "Pillow==12.2.0", "opencv-python-headless==4.10.0.84"],
+  "constraints": [
+    "numpy==1.26.4",
+    "scipy==1.12.0",
+    "scikit-learn==1.4.2",
+    "scikit-image==0.24.0",
+    "pandas==2.2.2"
+  ],
   "bundleRepo": "deepsafe/feature-bundles",
   "bundles": {
     "background-removal": {
@@ -408,15 +415,15 @@
       "archives": {
         "amd64-gpu": {
           "file": "v2.0.0/ocr-amd64-gpu.tar.gz",
-          "sha256": "80db475eb442899cedad3590b18f080225deaf50985224b302475fbfdf5883f6",
-          "compressedSize": 5927587552,
-          "extractedSize": 9367838720
+          "sha256": "2a00a3184f6a635f1fa9ae2a6517ad740a11f9e5ff58c098d2fd369a2bb1e16b",
+          "compressedSize": 5933794814,
+          "extractedSize": 9343914078
         },
         "arm64-cpu": {
           "file": "v2.0.0/ocr-arm64-cpu.tar.gz",
-          "sha256": "569ca2ddad50529c4887da494ca1704bbe719ddc0253a5eb76002a2d87812c60",
-          "compressedSize": 1974010780,
-          "extractedSize": 3092260263
+          "sha256": "6868c264069dcb74c6675c0b1f58dc1c9f60d9aa4459725e3dbde07a99a6a09a",
+          "compressedSize": 1984611355,
+          "extractedSize": 3103671712
         }
       },
       "packages": {

docker/verify-bundle.sh

@@ -171,7 +171,11 @@ case "${BUNDLE_ID}" in
     check_imports "torch onnxruntime mediapipe"
     ;;
   ocr)
-    check_imports "paddleocr paddle"
+    # scipy/scikit-learn ship inside this bundle via paddleocr's dependency
+    # closure. Importing them here (not just paddleocr) catches a numpy-2.x-ABI
+    # strand on the numpy==1.26.4 base; the "numpy.dtype size changed" class
+    # that a paddleocr-only import misses because paddle lazy-loads them.
+    check_imports "paddleocr paddle scipy sklearn"
     ;;
   transcription)
     check_imports "faster_whisper"