Chef's resources and resource collection and notifications scheme is entirely declarative. And after watching users beat their heads against Chef for a decade the thing that users really like is using declarative resources that other people wrote. The thing that they hate doing is trying to think declaratively themselves and write their own declarative resources or use the resource collection properly. People really want the glue code that they need to write to be imperative and simple.
The biggest issue that Chef had was the "two-pass parsing" design (build the entire resource collection, then execute the entire resource collection) along with the way that the resource collection and attributes were two enormous global variables which were mutable across the entire collection of recipe code which was being run, and then the design encouraged you to do that. And recipes were kind of a shit design since they weren't really like procedures or methods in a real programming language, but more like this gigantic concatenated 'main context' script. Local variables didn't bleed through so you got some isolation but attributes and the resource collection flowing through all of them as god-object global variables was horrible. Along with some people getting a bit too clever with Ruby and Chef internals.
I had dreams of freezing the entire node attribute tree after attribute file processing before executing resources to force the whole model into something more like a functional programming style of "here's all your immutable description of your data fed into your functional code of how to configure your system" but that would have been so much worse than Python 2.7-vs-3.0 and blown up the world.
Just looking at imperative-vs-declarative is way too simplistic of an analysis of what went wrong with Chef.
HCL is also not pure declarative code either. It can invoke non-declarative functions and can do loops based on environment variables, so in that sense there is really no difference between Pulumi and Terraform. The only real difference is that HCL is a terrible language compared to say Python.
I'm actually fairly sure HCL is Turing complete, it has loops and variables. But even if it is not all the way turing complete it's pretty close.
I have never seen Pulumi or CDKTF stuff work well. At some point are you simply writing a script and abandoning the advantages of a declarative approach
There are many graveyards filled with places that tried to start writing logic into their IaC back in the Chef/Puppet era and made a huge mess that was impossible to upgrade or maintain (recall that Chef is more imperative/procedural, whereas in Puppet you describe the desired end state). The Chef/Pulumi approach can work, but it requires one person who is draconian about style and maintenance. Otherwise, it turns into a pile of garbage very quick.
Terraform/Puppet's model is a lot more maintainable for longer terms with bigger teams. It's just a better default for discouraging patterns that necessitate an outsized investment to maintain. Yes HCL can be annoying and it feels freeing to use Python/TS/whatever, but pure declarative code prevents a lot of spaghetti.