// ── einsum — Einstein summation notation ──
//
// Supports the following common ML patterns via direct dispatch to
// optimised coeus-ops kernels:
//
// "ij,jk->ik" matrix multiply (→ matmul)
// "bij,bjk->bik" batched matmul (→ bmm)
// "bi,bj->bij" outer product per batch (→ broadcast + mul)
// "ij->ji" 2-D transpose (→ permute [1,0])
// "...ij->...ji" last-2D transpose (→ permute swap last two dims)
// "i,i->" dot product (→ element-wise mul + sum)
// "i,j->ij" outer product (→ unsqueeze + broadcast + mul)
// "ii->" trace (→ diagonal sum)
// "ij,j->i" matrix-vector (→ matmul with shape reshape)
// "bik,bk->bi" batched matvec (→ reshape + batched matmul + squeeze)
//
// Unrecognised patterns and subscript/shape mismatches are typed
// [`BackendError`](coeus_core::BackendError) failures, never panics.
//
// The module is split by pattern arity so each leaf owns one recognition
// table: [`dispatch`] parses the subscript and counts operands, [`single`] and
// [`binary`] hold the one- and two-operand arms, [`ternary`] the three-operand
// chain decomposition, and [`subscript`] the typed rejections they share.
pub use einsum;
pub use einsum3;