Step executor throws when a step error has a non-writable stack, so the step is never retried or failed #4369

@workflow/core 5.0.0-beta.55, dist/runtime/step-executor.js. Both the FatalError path and the retry path apply the normalized stack by assignment:

if (types.isNativeError(err) && normalizedStack) {
    err.stack = normalizedStack;
}

If the thrown error's stack is a non-writable data property, that assignment throws a TypeError inside the executor's catch block. As a result:

• The executor writes neither step_retrying nor step_failed.
• The queue handler logs Queue handler failed for message ... on delivery attempt 1; retrying in 1s: TypeError: Cannot assign to read only property 'stack' of object '...'.
• The redelivery finds the step already running, so it stays at attempt 1 until the execution lease expires. On our deployment that was about 14 minutes.

A transient step error becomes a long stall with no step event. Postgres.js errors on Node 22+ are one source of such errors (porsager/postgres#1227). Frozen errors and some third-party errors are others.

Repro

A step that throws:

const e = new Error('transient')
Object.defineProperty(e, 'stack', { value: e.stack, writable: false })
throw e

Expected: the step retries or fails, with the normalized stack in the event payload.
Actual: the queue handler fails and the step stays running until its lease expires.

Suggested fix

Don't mutate the caller's error. Either pass normalizedStack to the dehydrator, or guard the assignment (a try/catch, or Object.defineProperty when the property is configurable). In the published dist, the FatalError site is around line 1023 and the retry site around line 1131.