My first big problem is that you can't have more than one component per file, much like single file components in Vue. If you're working on a component that has many sub-components it becomes very tedious to switch around files. In Vue instead of creating new small components when using SFC I tended to complicate the template and logic of the current component... This made the workflow less tedious at first but I quickly learned is a terrible long-term approach. In more JS centric frameworks like React or Mithril I don't have this issue. It's ironic since Rich Harris preaches the "1 screen rule" and this completely breaks that.
The other issue I had with Svelte is the official reactive stores. The first weird thing is in Svelte a store is a reactive variable, not a bunch of reactive state like in other libs/frameworks. It quickly becomes tedious if you need to model semi complex data. Vue has the same problem in Vuex with its shallow (and dumb) objects. MobX IMO is so much better in terms of data modeling since you just define your models in any way you wish (pojos, classes, etc) and you just add decorators for the reactive properties/variables.
You can indeed add MobX to Svelte but then what's the point? It would be better to use Mobx-JSX which adds minimal overhead over MobX and is insanely fast.
https://github.com/ryansolid/mobx-jsx
Of course the problem with MobX is that while it has very nice dev ergonomics it's a very complex piece of machinery.
After fiddling with reactive state for about 4-5 years I've come to the conclusion the best approach is the one that Mithril and Imba follow: no reactivity. Your data is whatever you want and you don't need expensive tooling like MobX/Redux/Vuex to make it reactive. Whenever there is a user event everything is re-rendered automatically and the v-dom or memoized system take care of rendering the DOM parts that need to change. If the data is updated from an API response or some WebSocket the developer just calls a method to let know the renderer the data has changed. It might not be the ultimate fastest approach in terms of performance, but it makes development much more simple and it's fast enough for 99% of use cases.