[Code scan] Preserve documented 3x3 cell matrices instead of collapsing them to diagonals

[njzjz]

This issue is a Codex global repository scan finding for deepmodeling/reacnetgenerator at commit 4fa8e2b.

ReacNetGenerator.__init__() documents cell as accepting (3,3), (3,), or (9,) input, but the current normalization treats any object with len(cell) == 3 as a vector and passes it to np.diag().

Relevant code:

reacnetgenerator/reacnetgenerator/reacnetgen.py

Lines 275 to 282 in 4fa8e2b

	if self.cell is not None:
	if len(self.cell) == 9:
	self.cell = np.array(self.cell).reshape((3, 3))
	elif len(self.cell) == 3:
	self.cell = np.diag(self.cell)
	else:
	raise RuntimeError(
	"cell must be (3,3) array_like or (3,) array_like or (9,) array_like"

For a normal 3x3 matrix such as [[1,0,0],[0,2,0],[0,0,3]], len(cell) == 3, so line 279 produces the 1-D vector [1, 2, 3] instead of preserving the matrix. Downstream coordinate-based PBC detection expects matrix indexing, e.g. Open Babel setup reads cell[0][0], cell[0][1], etc.

Downstream use:

reacnetgenerator/reacnetgenerator/_detect.py

Lines 416 to 423 in 4fa8e2b

	-------
	list[list[int]]
	Connected atoms for each atom.
	list[list[int]]
	Bond orders for each atom. 12 is an aromatic bond.
	"""
	# Check if ASE mode is enabled
	if self.use_ase:

Impact:

• documented 3x3 cell inputs can be silently converted to the wrong shape
• PBC bond perception can crash or use invalid geometry
• users passing full triclinic or non-cubic cells through the Python API or CLI --cell can get incorrect behavior

Suggested fix:

Convert cell to a NumPy array first and branch on array.shape or array.size, preserving (3, 3), diagonalizing (3,), and reshaping flat (9,). Add regression tests for list-of-lists and flat 9-value cells.