Fix Julia 1.12: replace Cassette with code_typed-based branch detection

[ChrisRackauckas-Claude]

Problem

Julia 1.12 changed generated functions to require CodeInfo return values instead of Expr (see changelog). Cassette v0.3.14 does not implement this change, so every call to hasbranching fails on 1.12 with:

As of Julia 1.12, generated functions must return `CodeInfo`. See `Base.generated_body_to_codeinfo`.

This is an upstream Cassette limitation with no imminent fix.

Solution

Remove the Cassette dependency entirely and replace the IR-pass approach with a direct code_typed call:

function hasbranching(f, x...)
    is_leaf(f, x...) && return false
    results = code_typed(f, Tuple{Core.Typeof.(x)...}; optimize = false)
    isempty(results) && return false
    ci = first(results)[1]
    return any(isa(s, Core.GotoIfNot) for s in ci.code)
end

code_typed(...; optimize=false) gives type-inferred but unoptimized IR — exactly the same level Cassette's pass was inspecting. The semantics are preserved:

• x < 0 ? -x : x → GotoIfNot in IR → true
• ifelse(x < 0, -x, x) → function call, no GotoIfNot → false
• true ? 1 : 0 → GotoIfNot in unoptimized IR → true (constant folding only happens at optimize=true)
• x1, x2 = x destructuring → GotoIfNot stays inside iterate, invisible at the caller level → false

Breaking change: extension API

The Cassette.overdub specialization pattern is replaced by is_leaf dispatch:

# Old (broken on 1.12):
function FunctionProperties.Cassette.overdub(::FunctionProperties.HasBranchingCtx, ::typeof(f), x...)
    f(x...)
end

# New:
FunctionProperties.is_leaf(::typeof(f)) = true

Changes

• src/FunctionProperties.jl: 170 lines → 45 lines; no deps
• Project.toml: drop Cassette and DiffRules deps; bump to v0.1.4
• test/runtests.jl: update extension example to use is_leaf

All tests pass on Julia 1.10 and 1.12.

Note: Please ignore until reviewed by @ChrisRackauckas.

[topolarity]

@ChrisRackauckas FYI I took a peek since you mentioned this in the meeting today

It looks like this PR only checks the immediate function and doesn't consider any callees - I assume that's a problem for correctness

[ChrisRackauckas]

Oh good point, since it needs to check all functions that aren't inlined. Okay, that's not too hard though, and can still avoid doing a full cassette approach.