[2.0 / next] A memo reading latest(a) beside b() computes a torn pass on a batched write; if that pass throws, the error sticks #4365

Found by the async app generator (packages/web/test/async-gen, branch claude/project-thread-bgafe5), minimized. There is no async anywhere in this case.

const [a, setA] = createSignal(1);
const [b, setB] = createSignal(0);
render(() => {
  const m = createMemo(() => {
    const v = latest(a) + b() + 1;
    if (v % 4 === 3) throw new Error("m:" + v);
    return v;
  });
  return <Errored fallback={e => `[E ${e()?.message}]`}><span title={String(m())} /></Errored>;
}, el);
flush();
setB(3); setA(2); flush();
// m computes latest(a)=2 with b()=0 → 3 → throws, and never re-runs: screen stays [E m:3]

Expected: m = 2 + 3 + 1 = 6, <span title="6">. A fresh render with a=2 and b=3 shows that.

Actual on next (1f558e5): the memo runs one pass with the new latest(a) beside the old b(). When that pass throws, the boundary keeps the error, and nothing re-runs m with b=3.

Without the throw, the torn pass still reaches effects. In a memo returning [latest(s), s()] after set(2), a render effect commits [2,1] before [2,2]. That is a torn tuple from a plain write with no transition held, which "latest tears" (#4284) doesn't seem to cover.