1 Commits (fix-derived-leak-on-throw)

Author	SHA1	Message	Date
Nic Polumeyv	f2648b3537	fix: end a restored reaction context at the end of its synchronous segment (#18694) ... When an async expression resumes after a pickled `await`, the thunk returned by `save()` in `reactivity/async.js` calls `restore()` to re-arm `active_reaction` for the rest of the expression, then disarms it with `queue_micro_task(unset_context)`. Any microtask already queued before that one runs inside the restored context. If it writes to a source, `set()` throws `state_unsafe_mutation` in production, since the guard is not dev-only. #18453 introduced the queued disarm and noted this case in review as unavoidable. SvelteKit hits it in practice: its fetch continuations write to internal `$state` (sveltejs/kit#16914), and a user's `$derived((await q()).length)` resuming in the same tick makes that write throw and drops the update signal. The context restored by a `save` thunk now ends with the synchronous segment it was restored in. A `restored` flag is set by the thunk and consumed on entry to `save` and `track_reactivity_loss`, so every suspension ends it; once an expression contains a pickled await, the analysis pickles every later await in it too (`has_pickled_await` on `ExpressionMetadata`), so a trailing await compiles to `$.save` rather than a bare `await`. At the end of the body, `async_thunk` in `3-transform/client/utils.js` wraps the return expression in `$.unsave(...)` when the metadata has a pickled await. If the body throws instead, the context is unset by `async_derived`'s existing `finally`, as before. The queued microtask in `save` is removed. Output is unchanged for expressions that pickle nothing (`$derived(await a)` compiles byte for byte the same). Expressions with a pickled await gain one `$.unsave(` call per body, and their trailing await becomes a `$.save`, 4 to 6 bytes gzipped in the added tests. At runtime a boolean write replaces a queued microtask per resume. `bench:compare` shows no difference outside run-to-run noise. Two runtime tests reproduce the throw without any library involved, one in dev and one with the prod `await` shape, and fail on `main`. --------- Co-authored-by: Simon Holthausen <simon.holthausen@vercel.com>	16 hours ago