Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 100.00%. Comparing base (ba7da77) to head (3faf156).
Report is 6 commits behind head on main.

@@            Coverage Diff            @@
##              main      #224   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           15        15           
  Lines         1758      1816   +58     
=========================================
+ Hits          1758      1816   +58     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

Todo:

• Set it as an option inside the algorithms
• Fill with zeros inside the loop

@amontoison moving this from "enabled by default" to "enabled on demand" took a bit more work than expected, and I had to add a lot more tests to be confident. I wouldn't mind another round of review

@gdalle
I think it's fine to not add a kw allow_denser after thinking about it.
It introduces additional complexity for the user that we could avoid that if we are careful in decompress!.
We can do something like this by default to protect the users that provides sparse matrices with a sparsity pattern that doesn't match the one used for coloring.

function decompress!(A::SparseMatrixCSC, B::AbstractMatrix, result::ColumnColoringResult)
    (; compressed_indices) = result
    S = result.bg.S2
    nzA = nonzeros(A)
    if nnz(A) == nnz(S)
        for k in eachindex(compressed_indices)
            nzA[k] = B[compressed_indices[k]]
        end
    elseif nnz(A) > nnz(S)
        rvA, rvS = rowvals(A), rowvals(S)
        shift = 0
        for j in axes(S, 2)
            for k in nzrange(S, j)
                i = rvS[k]
                found = false
                while (k + shift) < A.colptr[j+1] && !found
                    if rvA[k + shift] == i
                        found = true
                        nzA[k + shift] = B[compressed_indices[k]]
                    else
                        shift += 1
                        nzA[k + shift] = zero(eltype(A))
                    end
                end
                !found && error("...")
            end
        end
        nzA[(nnz(S) + shift + 1):end] .= zero(eltype(A))
    else
        error("...")
    end
    return A
end

Unfortunately I think your version is incorrect: it doesn't account for the fact that we could find a coefficient (j', i) inside A with j' < j, inside a column that is for instance empty in S. In that case, your while !found loop will stop too early.
As discussed, I suggest we keep my version for now and improve later on, since this is just about error messages being nicer.

Following the discussion in SciML/OrdinaryDiffEq.jl#2653 (comment), this is put on hold for the time being. Let's not delete the branch in case we need it again.