Repeated index in `A[I, J] = B` corrupts the matrix (duplicate row) or drops values (duplicate column)

[yurivish]

When assigning to a repeated index, a SparseMatrixCSC instead stores multiple assigned values. The last write should win, but with a sparse matrix the assignment corrupts the matrix:

julia> using SparseArrays

julia> A = spzeros(2, 2);

julia> A[[1, 1], 1] = [5.0, 6.0];

julia> Matrix(A) 
2×2 Matrix{Float64}:
 6.0  0.0
 0.0  0.0

julia> A # incorrect value and number of stored entries
2×2 SparseMatrixCSC{Float64, Int64} with 2 stored entries:
 5.0   ⋅
  ⋅    ⋅

julia> sum(A) # incorrect sum
11.0

julia> sum(Matrix(A))
6.0

julia> B = zeros(2, 2);

julia> B = zeros(2, 2);^C

julia> B[[1, 1], 1] = [5.0, 6.0];

julia> B
2×2 Matrix{Float64}:
 6.0  0.0
 0.0  0.0

Possible cause (AI-diagnosed)

setindex!(A::AbstractSparseMatrixCSC, V::AbstractVecOrMat, I, J) in sparsematrix.jl sorts I and J but assumes they contain no repeats. In the column merge

rowB = I[rowvalB[ptrB]]          # two rows of B map to the same row of A: both are emitted
...
colptrA[col+1] = ptrS
colB += 1                        # one column of B per distinct column of A: later duplicates in J are never visited

and the final _checkbuffers(A) only checks buffer lengths. A minimal fix is to fall back to the generic element-wise loop when !allunique(I) || !allunique(J) (which gives the Matrix semantics), or to throw an ArgumentError for repeated indices.