$derived on a class instance returns a stale value after the effect that incidentally owned it at construction is destroyed

[reinhard-sanz]

Describe the bug

A $derived declared as a class field is "owned" by whatever effect happens to be active when the instance is constructed (derived.parent = active_effect). If that instance outlives the effect, e.g. a long-lived model object that was new'd inside a component which later unmounts, the derived is frozen to its last value and never recomputes again, even though its $state dependencies keep changing.

This is a regression introduced in 5.55.3 by #17921; pinning to 5.55.2 restores correct behavior.

The trap is that the owning effect is non-deterministic and semantically meaningless for an object whose lifetime isn't tied to any component, it's just whichever effect was on the stack at new-time.

Reproduction

https://svelte.dev/playground/86e9f769bb1d4021af59f2d289b73448?version=5.56.4

Click the button twice.
Expected: after step 2, box.value === 'C' so box.doubled === 'CC'.
Actual (5.55.3+): box.doubled === 'BB', frozen at the value from before the child unmounted. Reading it inside a reactive context (template/$effect) instead reconnects and unfreezes it, which makes the staleness order-dependent and hard to spot.

Logs

System Info

System:
    OS: macOS 26.4.1
    CPU: (14) arm64 Apple M4 Pro
    Memory: 3.37 GB / 48.00 GB
    Shell: 5.9 - /bin/zsh
  Binaries:
    Node: 22.18.0 - /Users/reinhard/.nvm/versions/node/v22.18.0/bin/node
    npm: 10.9.4 - /Users/reinhard/.nvm/versions/node/v22.18.0/bin/npm
    bun: 1.3.14 - /Users/reinhard/.bun/bin/bun
  Browsers:
    Brave Browser: 146.1.88.134
    Chrome: 149.0.7827.156
    Firefox: 152.0.3
    Safari: 26.4
  npmPackages:
    rollup: ^4.56.0 => 4.56.0

Severity

blocking an upgrade

[reinhard-sanz]

I think this is a small extension of a fix that already landed, not a new edge case.

Same root cause as #18139 (another #17921 regression), and still reproduces on the latest 5.56.4.

The mechanism is in execute_derived:

if (
  !is_destroying_effect &&
  parent !== null &&
  derived.v !== UNINITIALIZED &&
  (parent.f & (DESTROYED | INERT)) !== 0
) {
  w.derived_inert();
  return derived.v; // ← stale frozen value
}

A class-field $derived on a long-lived object is "owned" by whatever effect was on the stack at new-time. When that effect is destroyed, reads outside a reactive context hit this branch and return the stale derived.v (with a derived_inert warning) even though its $state deps have changed.

Crucially, the derived.v !== UNINITIALIZED carve-out here was added in #18228 to fix #18139, with the rationale: "there are a few unavoidable cases where we have to execute the derived even if we otherwise wouldn't, because of its lazy nature." A dirty derived on an object whose lifetime isn't tied to any component is exactly another such case, the owning effect is semantically meaningless for the derived's validity.