Store.notify() drops updates that arrive while a key's observers are executing

[owahltinez]

Summary

Store.notify() discards a notification that arrives while that key's observers are running, instead of remembering that the value changed again. The DOM is then left permanently disagreeing with the store: the last write a key received never reaches its observers, and nothing retries.

src/store.ts:

const current = this._notify.get(key);
// Skip if observers are already executing for this key (prevents infinite loops).
if (current === "executing") return;

The comment names the real requirement — an observer that writes its own key must not recurse forever — but dropping is a stronger remedy than that requires, and it silently loses legitimate updates from other writers.

Measured

Observed on a page that writes one store key from pointermove and binds it to one element's style. A MutationObserver on that attribute, counted against the pointermove events the page received, over a single gesture:

	pointermoves	style writes landed	lost
Chromium 412x915	41	38	3 (7%)
Firefox 412x915	41	37	4 (10%)
Chromium, CPU throttled 20x	41	32	9 (22%)

Every landed write carried a distinct value (new Set(writes).size === writes.length in all runs), so these were not de-duplicated no-ops — they were dropped. The loss scales with how long observers take, which is what makes it a mobile-first symptom: on a phone the same code loses most of its updates.

The visible bug in our case: an element revealed by one key while the key carrying its position was dropped, so it painted at its untransformed origin.

Suggested fix

Coalesce rather than drop — latest-wins, at most one extra pass:

if (current === "executing") {
  this._notifyAgain.add(key);   // or store "executing-dirty" in the same map
  return;
}

and in the finally after await Promise.all(...), if the key was marked, schedule one more notification. That preserves the recursion guard (a self-writing observer converges after one extra pass instead of looping) while making a lost update impossible.

Happy to send a PR with a regression test if the approach looks right.

Version

mancha@0.22.6; the same code is on master.