David Widmann

> These last two points can be eliminated with constrained stepping though. If constrained stepping still is unstable, then it is likely just a property of the Rosenbrock integrator. I...

> We should also check with autodiff=false. The algorithms are unstable with finite differencing as well: ```julia sol = solve(prob_dde_rober2, MethodOfSteps(Rosenbrock23(; autodiff = false)); initial_order = 1) ``` ...

> Can you run an SDIRK like TRBDF2 and see if that gives the same results as Rosenbrock? Unfortunately, using TRBDF2 does not help: ```julia sol = solve(prob_dde_rober2, MethodOfSteps(TRBDF2()); initial_order...

Hmm it's really strange that we can not verify the instability around 16.8.

I ran RADAR5 on this problem with a delay of 0.03 (which should explode according to Hairer and Guglielmi) and could compute a reasonable solution on the whole interval [0.0,...

That might very well be the case... I already assumed something along those lines when I got `NaN` values in the first runs above.

> seems reasonable to make `--color=yes` a default flag I would say Initially, I thought about including a fix for this issue in #73. But I started to question whether...

Can we have both? That is, could we forward `f(arg1, ::Type{TV}, ...)` to `f(arg1, ::TypeWrap{TV}, ...)`?

I'd also make it a breaking release.

> Maybe we can push this a bit further by introducing a block syntax for `:=`: If we consider a block syntax, maybe something like ```julia julia> @model function demo()...