[devmotion]

At a Hermitian A, the eigen/eigvals frules and rrules for StridedMatrix delegated to the rules for Hermitian(A). Their (co)tangents are those of x -> eigen(Hermitian(x)), which reads only the upper triangle of x. That is wrong for unstructured matrices: the rrules returned upper-triangular cotangents and the frules ignored the lower triangle of ΔA (FluxML/Zygote.jl#1369).

This PR:

• Eigenvalues: tangents uᵢ' ΔA uᵢ and cotangents U Diagonal(Δλ) U', with the type of the primal eigenvalues (so they are only projected onto the reals when the primal eigenvalues are real-typed).
• Eigenvectors: at Hermitian A, eigen uses the Hermitian solver, whose eigenvectors follow a different phase convention than those of non-Hermitian matrices. They are therefore only differentiable along Hermitian directions, so the rules use the Hermitian part of V' ΔA V for the eigenvector (co)tangents (and the Hermitian phase convention).
• Both cases now share the code path of the generic rules (V' instead of V \ since V is unitary). The special handling of sortby is no longer needed: the rules call eigen!(A; kwargs...) itself, which sorts exactly as the primal does, and the (co)tangents are computed per eigenpair from the sorted result. _sorteig!_fwd became unused and is removed.
• Fixes the failures on Julia 1.14 (pre): LinearAlgebra.jl#1717 made eigenvalues of complex matrices complex-typed also on the Hermitian path, so the old rules returned real tangents for complex primals. It also exposed a Core.Box in the eigen pullback (self-recursive thunk method), which made the eigvals rrule non-inferable; the pullback is now a single method.
• Tests compare against finite differences of the unwrapped eigvals (whenever the eigenvalue type admits all directions) and of eigen ∘ hermitianpart for eigenvectors, with non-Hermitian ΔA, instead of x -> eigen(Matrix(Hermitian(x))). Includes a regression test for Zygote#1369 (∇ sum(eigvals(A)) == I), tests that the (co)tangents follow the order of the eigenpairs for a custom sortby, and that the imaginary parts of cotangents of real-typed eigenvalues are ignored.

Supersedes #841 (thanks @CarloLucibello for the analysis there), which fixed the values-only rrule for real matrices.

CI passes on Julia lts, 1 (incl. x86) and pre, as well as the Zygote and Yota integration tests.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.64%. Comparing base (e996d60) to head (550778b).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #853      +/-   ##
==========================================
+ Coverage   96.57%   96.64%   +0.07%     
==========================================
  Files          28       28              
  Lines        3616     3603      -13     
==========================================
- Hits         3492     3482      -10     
+ Misses        124      121       -3     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.