[huashengshao]

Color-pair overlaps do not define operator matrices in a non-orthonormal or redundant color basis. Multiplying those overlaps loses the ordering needed for overlapping color insertions.

This adds ColorDipole and ColorDipoleProduct, which act on exact retained color tensors. Each dipole reuses two physical-charge emissions and contracts their adjoint indices with the existing network reducer. Ordered products apply right-to-left on the ket, preserving shared-leg order, complex phases, symbolic Nc powers and inverse-square-root normalization. There is no Gram inversion or floating conversion.

The supported representation class is the existing all-outgoing fundamental, antifundamental and adjoint color space. Unknown or singlet insertion labels, mismatched typed spaces and normalization overflow fail explicitly. More representations require corresponding emission and reduction rules. Production logic does not select any process, flavor or multiplicity.

Validation: all 27 tests matching color:: pass, including seven new regression groups checking independent pair overlaps, exact Casimirs and color conservation, noncommuting overlapping products, directed phases and normalization, invalid/extreme labels and crossing. The tripole test retains F = f_abc T1^a T2^b T3^c = i[P12,P23]; it checks the explicit literal-to-network trace orientation and independently contracts exact Pauli matrices to obtain the nonzero SU(2) off-diagonal values ±3i/2.

The test binary ran with PATH=/nonexistent and one test thread after an offline locked build with default features plus symbolica-runtime,python-generation-bridge,loop-v2, using Symbolica/Numerica 9cdc2949cdb153478c1eca0d40690dc0abfc616e. The other 1,159 tests were filtered out. Source and corrected tripole-oracle review are independent of the implementation.

Stacked on #17 to share its provider baseline; this PR changes only the color API and its tests. It does not yet connect these operators to physical amplitude evaluation or establish NNLO predictions.