Turing.jl identical likelihood but different `MAP()` result?

Specific Domains Probabilistic programming
, ,
1

sample script:
https://gist.github.com/Moelf/ab2f10ce45873bb91eaec60ace8388c0

you will need

]dev https://github.com/JuliaHEP/LiteHF.jl
]add Turing, Optim

to compare, flip this line:
https://github.com/JuliaHEP/LiteHF.jl/blob/c68a70da2c54a48424683cf8ac538e0e67e762c7/src/modifiers.jl#L16

between InterpCode4 and InterpCode0.

I see:

ModeResult with maximized lp of -13.22
3-element Named Vector{Float64}
A               │ 
────────────────┼───────────
:μ              │    1.29259
Symbol("θs[1]") │ -0.0411768
Symbol("θs[2]") │  0.0178776

julia> loglikelihood(mymodel, (μ=0, θs=[1,0]))
-10.662623923767459

and

ModeResult with maximized lp of -13.22
3-element Named Vector{Float64}
A               │ 
────────────────┼──────────
:μ              │   1.30279
Symbol("θs[1]") │ -0.060553
Symbol("θs[2]") │ 0.0187052

julia> loglikelihood(mymodel, (μ=0, θs=[1,0]))
-10.662623923767459