[jverzani]

Problem

@syms c() a; Dict(c(0) => a) segfaults, as does hashing any Sym (hash(a), Set, unique).

hash is defined for Sym, but it called PyObject_Hash directly through a raw ccall. PyCall 1.97 releases the GIL between calls and wraps its own C-API calls in @with_GIL. The raw call ran without the GIL and Python crashed in slot_tp_hash.

Fix

• Defer to PyCall's hash(::PyObject, ::UInt), which acquires the GIL. The old method was a workaround until PyCall shipped a 2-arg hash, and 1.97 has one.
• Bump PyCall compat to 1.97 (1.96.x only defines the 1-arg hash).
• Add a regression testset, run before the shared SymPyCore suite.
• Version 2.3.4. Also ignores JetBrains IDE files.
• CI: replace deprecated actions/cache@v1 (now hard-failing) in ci-nightly.yml with julia-actions/cache, bump all actions to current majors, and add .github/dependabot.yml for GitHub Actions and Julia deps.

CI note

On Julia 1.13 (the 1 job), two test-legacy.jl tests in SymPyCore fail independently of this PR. They expect 3 ^ [x 1] and 3 ^ [x 1; 1 x] to throw a MethodError. In 1.13, LinearAlgebra's ^(::Number, ::AbstractMatrix) calls exp(A) instead of exp!, so the calls no longer throw. The fix belongs in SymPyCore, and this package's compat will pick up that release. The hash testset runs before the shared suite, so it still reports on 1.13.

🤖 Generated with Claude Code