Can't sum partially initialized non-isbits Hermitian or Symmetric matrix with Diagonal or SymTridiagonal

[araujoms]

using LinearAlgebra
a = Matrix{BigFloat}(undef, 2,2)
a[1] = 1; a[3] = 1; a[4] = 1
a = Hermitian(a)
b = Symmetric(a)
c = Diagonal([1,1])
d = SymTridiagonal([1 1;1 1])
a+c
a+d
b+c
b+d

I've also tested with UpperTriangular, LowerTriangular, UnitUpperTriangular, UnitLowerTriangular, they work fine.

I've noticed this issue when testing my package with Julia 1.11. It showed up because when I do

using LinearAlgebra
a = Hermitian(BigFloat.([1 1;1 1]))
b = a + a

I get a partially initialized b, unlike in previous versions.

[mbauman]

So an MWE for the regressed case is:

julia> using LinearAlgebra

julia> (2*Hermitian(BigFloat.([1 1;1 1]))) + Diagonal([1,1])
ERROR: UndefRefError: access to undefined reference

This is probably from JuliaLang/julia#54473.

[araujoms]

Indeed, that's a regression, but the cause is JuliaLang/julia#52942.

[jishnub]

These seem to fail on v1.10.5 as well

[araujoms]

The regressed case is not the first code I posted, but instead

using LinearAlgebra
(2*Hermitian(BigFloat.([1 1;1 1]))) + Diagonal([1,1])
(2*Hermitian(BigFloat.([1 1;1 1]))) + SymTridiagonal([1,1],[1])

Just tested again on 1.10.5, it succeeds.