fix: separability for nested CompoundModels (swev-id: astropy__astropy-12907)

[casey-brooks]

Description

• add regression coverage for CompoundModel nesting
• fix _cstack to concatenate right coord matrices without dropping structure
• commits include [skip ci] per instructions to avoid running CI

Reproduction Steps

1. Plain conjunction (matches expectation):

   from astropy.modeling import models
   from astropy.modeling.separable import separability_matrix
   
   compound = models.Linear1D(10) & models.Linear1D(5)
   print(separability_matrix(compound))
   # expected / observed
   # [[ True False]
   #  [False  True]]

2. Mixed compound (non-nested) remains correct:

   compound = models.Pix2Sky_TAN() & models.Linear1D(10) & models.Linear1D(5)
   print(separability_matrix(compound))
   # expected / observed
   # [[ True  True False False]
   #  [ True  True False False]
   #  [False False  True False]
   #  [False False False  True]]

3. Nested compound (pre-fix observed incorrect duplication):

   inner = models.Linear1D(10) & models.Linear1D(5)
   compound = models.Pix2Sky_TAN() & inner
   print(separability_matrix(compound))
   # expected
   # [[ True  True False False]
   #  [ True  True False False]
   #  [False False  True False]
   #  [False False False  True]]
   # observed (before fix)
   # [[ True  True False False]
   #  [ True  True False False]
   #  [False False  True  True]
   #  [False False  True  True]]

Failing Pytest Trace (pre-fix)

______________________ test_separability_nested_compound _______________________

    def test_separability_nested_compound():
        inner = models.Identity(2)
        compound = models.Identity(2) & inner
        expected = np.eye(4, dtype=bool)
>       assert np.array_equal(separability_matrix(compound), expected)
E       assert False
E        +  where False = <function array_equal at 0xffff98d088b0>(array([[ True, False, False, False],
E       [False,  True, False, False],
E       [False, False,  True,  True],
E       [False, False,  True,  True]]), array([[ True, False, False, False],
E       [False,  True, False, False],
E       [False, False,  True, False],
E       [False, False, False,  True]]))
E        +    where <function array_equal at 0xffff98d088b0> = np.array_equal
E        +    and   array([[ True, False, False, False],
E       [False,  True, False, False],
E       [False, False,  True,  True],
E       [False, False,  True,  True]]) = separability_matrix(<CompoundModel()>)

../modeling/tests/test_separable_regressions.py:28: AssertionError

Fixes #66

[casey-brooks]

Local Validation

• python3 -m pytest test_separable_regressions.py
  • 4 passed, 0 failed, 0 skipped

[rowan-stein]

Requesting review:

This PR fixes separability_matrix for nested CompoundModels by preserving the right-hand nested structure in _cstack.

• References Issue Modeling's separability_matrix incorrect for nested CompoundModels #66
• Includes reproduction steps and failing pytest stack trace captured pre-fix
• Adds regression tests for plain &, mixed compound, nested compound, and deeper nesting equivalence
• Commits are marked with [skip ci] per instructions (CI intentionally not run)

Focus areas:

• Correctness of _cstack bottom-right block assignment
• Test coverage adequacy and clarity
• Consistency with existing modeling/separable conventions

[noa-lucent]

The fix swaps the all-ones placeholder in _cstack for the actual right-hand coord matrix, so nested CompoundModels retain their separability structure. The new regression suite exercises the plain, mixed, nested, and deeply-nested cases and matches modeling semantics (fails prior to the fix, passes here). Looks great—thanks!