Preview: https://svelte-dev-git-preview-svelte-16724-svelte.vercel.app/

Playground

pnpm add https://pkg.pr.new/svelte@16724

I feel a bit weird about adding a less-cryptic error for when matched is empty, but still not erroring at all if there are duplicate keys that don't result in an error. Is there some other proxy, besides matched being empty, for those other cases?

I also realised as a result of this PR that validate_each_keys doesn't always work, because the render_effect could run after the block update — for the test case in this PR, it fails with a is undefined. We could change that by turning the render_effect in validate_each_keys to a block, but maybe it would be better to move the check into reconcile inside an if (DEV) block?

Checking other cases requires computations. We cannot even do array.length === state.items.size - to_destroy.length (key duplication often results into a shortened list) because state.items may still contain transiting items from previous runs. Probably the simplest computation is a set of seen keys in the reconcile - validate keys during reconciling. (削除) Another place is inside create_item. (削除ここまで) If we want to add any computation to prod.

I vote for merging this. I just ran into this and it's really annoying to debug. Yes it's not going to catch all cases but I'd rather fail with a runtime error that gives me a hint at what goes wrong compared to a cryptic "undefined" runtime error.

I tried to move validate_each_keys inside reconcile but some tests failed, and I didn't get time to dig into why.

I moved validate_each_keys inside #each's block because, if thrown in reconcile, the error isn't caught by the boundary in async mode. Guess it is because reconcile is called in batch callbacks.

Btw, in prod the error isn't caught by boundary. I need help here.

Added the error emitting for the non-crashing case.
(削除) It doesn't emit the error at hydration in prod, but I guess it's fine. (削除ここまで)

Added ensuring the thrown error is caught by a boundary - is this the right approach?
I removed the tests because DEV is always true and thus we cannot really test the prod checks.
No idea why the TSGo test decided to die on an unrelated file.