fix(planner): replace insecure pickle artifacts with path-confined, validated NPZ loading

[mozluk]

Description

This PR addresses high-severity input validation and deserialization safety vulnerabilities within the PCT_planner repository[cite: 11]. It completely migrates the tomogram artifact system from vulnerable executable .pickle files to secure, non-object NumPy .npz archives, effectively neutralizing arbitrary code execution and path traversal risks[cite: 11].

Key Changes

• Safe Artifact Generation (tomography/scripts/tomography.py):
  • Changed the tomogram export format from pickle to compressed NumPy archives (.npz)[cite: 11].
  • The exportTomogram function now explicitly uses np.savez_compressed to package the required data structures safely[cite: 21].
• Strict Loading & Path Confinement (planner/scripts/planner_wrapper.py):
  • Replaced the vulnerable pickle reader with np.load(..., allow_pickle=False) to strictly disable NumPy object loading[cite: 11, 20].
  • Introduced the _resolve_tomogram_path helper to enforce path confinement, ensuring that caller-controlled filenames cannot break out of the configured tomo_dir via path traversal attacks[cite: 11, 20].
  • Added strict schema validation upon load to verify that all required keys (data, resolution, center, slice_h0, slice_dh) are present and that the tensor shape is exactly as expected[cite: 11, 20].
• Documentation & Tooling (planner/scripts/plan_global.py, README.md, .gitignore):
  • Updated CLI arguments in plan_global.py to specify that tomogram filenames should be provided without the .npz extension[cite: 22].
  • Updated README.md to document the new .npz export format and explicitly state that legacy pickle artifacts are intentionally rejected and must be regenerated[cite: 11, 23].
  • Added *.npz to .gitignore to prevent tracking generated artifacts, while retaining the legacy *.pickle exclusion[cite: 11, 24].

Action Required

• Existing tomogram .pickle artifacts are incompatible with this update and must be regenerated as .npz files[cite: 11, 23].
• A human reviewer must verify the native NumPy runtime execution locally, as the test environment lacked the required numpy dependencies to perform a full integration round-trip[cite: 11].