Loss of type stability in 1.12 for `group_varnames_by_symbol`

[penelopeysm]

using DynamicPPL, Test
vns_tuple = (@varname(x), @varname(y[1]), @varname(x.a), @varname(z[15]), @varname(y[2]))
@inferred DynamicPPL.group_varnames_by_symbol(vns_tuple)

This passes on 1.11 but fails on 1.12.

This test is also in our test suite so I'm completely stumped as to why CI hasn't been failing. I discovered this when running the test suite locally.

I compared the Manifest of my local test environment vs the one from this job and it's exactly the same.

DynamicPPL.jl/test/utils.jl

Lines 198 to 211 in 2020741

	@testset "group_varnames_by_symbol" begin
	vns_tuple = (
	@varname(x), @varname(y[1]), @varname(x.a), @varname(z[15]), @varname(y[2])
	)
	vns_vec = collect(vns_tuple)
	vns_nt = (;
	x=[@varname(x), @varname(x.a)],
	y=[@varname(y[1]), @varname(y[2])],
	z=[@varname(z[15])],
	)
	vns_vec_single_symbol = [@varname(x.a), @varname(x.b), @varname(x[1])]
	@inferred DynamicPPL.group_varnames_by_symbol(vns_tuple)
	@test DynamicPPL.group_varnames_by_symbol(vns_tuple) == vns_nt
	end

[mhauru]

Pkg.test(; coverage=true) hits this test failure, Pkg.test() doesn't. EDIT: Sorry, the other way around, Pkg.test(; coverage=true) doesn't fail. Trying to understand why.

[mhauru]

Reported this over at Pkg.jl, see above link.

[mhauru]

If this doesn't get fixed quickly on 1.12.2, we might have to put in a workaround. This is probably causing some performance regression somewhere, just not in our test suite as run on CI.

[penelopeysm]

Can't we just tell all users to always run Julia with coverage?