fix(mv): is_blade() now correctly handles null vectors and null blades

[utiberious]

Summary

• is_blade() previously delegated entirely to is_versor(), which returns False for null multivectors (no inverse → not a versor)
• This caused grade-1 null vectors like e0+e1 in G(1,1) and grade-2 null blades like (e0+e1)^e2 in G(1,2) to be incorrectly reported as non-blades
• Blade-ness is a metric-free concept (depends only on the outer product)

Fix

1. Not grade-homogeneous → False
2. Grade ≤ 1 → always True (scalars and vectors are always blades)
3. Non-null (invertible) → is_versor() path (unchanged)
4. Null case → outer-product squaring test: B ^ B == 0 (necessary for any blade; sufficient for grade 2)

The docstring notes the known limitation: for grades ≥ 3 in algebras of sufficiently high dimension, B ^ B == 0 may give false positives; full factorizability is not yet implemented.

Test

Added test_is_blade_null covering:

• Null 1-vector in G(1,1)
• Non-null 1-vectors
• Null 2-blade in G(1,2) with null constituent vector
• Mixed (non-grade-homogeneous) multivector

Closes #537

[utensil]

Thanks for the fix. Two issues to address before merging.

The not self.is_zero() guard lives in step 4 but the zero scalar never reaches it: i_grade = 0 is caught by i_grade <= 1 at step 2, returning True. Current behavior is False for mv(0).is_blade(). If zero should stay excluded (standard convention), the guard needs to be before step 2.

reflect_in_blade and project_in_blade gate on is_blade() then immediately compute / blade.qform(). For null blades, qform() is zero, so they'll divide by zero after this fix. Before, is_blade() = False for null blades effectively protected those methods. They need a null-blade guard now.

Also the test comment "null → no inverse → not a versor" is slightly off — galgebra's is_versor() checks V*V.rev() = 0, not the existence of an inverse. Minor but worth fixing.

[utensil]

On the grade ≥ 3 limitation: the docstring says "sufficiently high dimension" but it's worth being concrete. The standard counterexample in R⁶ (Dorst, Fontijne & Mann, Geometric Algebra for Computer Science, 2007):

B = v₁∧v₂∧v₃ + v₁∧v₄∧v₅ + v₂∧v₄∧v₆ + v₃∧v₅∧v₆

(B^B).is_zero() returns True and B.is_zero() is False, but B cannot be factored as u∧v∧w for any vectors u, v, w — so is_blade() would silently return True. A test pinning this known false positive (asserting is_blade() == True with a comment that it's a known limitation) would make the boundary explicit rather than prose-only.

[utensil]

Thanks for addressing the feedback. Zero guard is correctly placed, reflect/project guards look good, and the test comment is accurate now.