Nicely done. I still feel there're some bad code smells here. Let me illustrate in a separate issue without blocking this PR.

Originally posted by @hyf0 in #6860 (review)

Some design orientations

• All the rust crates (except rolldown_binding) should be written as if they were gonna be used as a rust libaray.
  • This enforces us to deisgn clean and elegant API and well-organized code structures, which makes us easily to maintain the code.
• There are cases that rules don't apply, but you should try to follow them as much as possible.

Validation like #6860 should be done within the crate rolldown's scope instead of being in rolldown_binding.

And why the PR can't do that. That's becuase Watcher is designed to receive the bundler from outside, so you can't do a advance check with NormalizedBundlerOptions and NormalizedWatcherOptions.

And why the watcher tries to receive the bundler from outside? That's becuase, in the current code, we only provide abstriction about BindingOptions -> BindingBundler, so binding watcher would have use this abstraction and then take the Bundler out of BindingBundler. (This's a classic example of not considering about the overall code design but just to make it run)

The practic way should be:

• We provide abstraction for BindingOptions -> BundlerOptions + Plugins
• BindingWatcher convert BindingOptions to BundlerOptions + Plugins and pass them to Wacher
• Watcher receive BundlerOptions + Plugins and perform validation inside of Watcher itself
• If something wrong, Watcher::new will return BuildErrors, and binding watcher will be responsible to convert binding errors to napi errors.

Misc

• binding_dev_engine has this problem too.
• NormalizedWatcherOptions doesn't exist, but you should get the gist of what I said.