Bug: Durable Object alarm firing under durableObjectStorage = (inMemory = void) crashes the workerd process #7191

Summary

Any Durable Object alarm that FIRES while the worker is configured with
durableObjectStorage = (inMemory = void) terminates the entire workerd
process with an uncaught KJ exception. The alarm handler body is irrelevant
(an empty handler crashes identically).

Local-dev configuration surface; production DOs are disk-backed.

Environment

• workerd 1.20260826.1, macOS arm64 (release build) and linux arm64 (ASAN build) — both reproduce identically
• compatibilityDate: 2026-08-26 and 2024-09-23 (both reproduce)

Repro

1. Config: one service, one DO class, durableObjectStorage = (inMemory = void).
2. Worker arms an alarm: await this.ctx.storage.setAlarm(Date.now() + 40);
3. Alarm handler: may be completely empty (return;).
4. When the alarm fires, workerd exits:

*** Fatal uncaught kj::Exception: kj/async.c++:2202: failed: Promise callback destroyed itself.

Discrimination

• With durableObjectStorage = (localDisk = "<writable disk service>") the same alarm fires and completes cleanly.
• deleteAlarm() before the scheduled time does not crash.
• ctx.abort() inside the handler does not crash (the DO resets instead); with the default options no retry was observed post-reset.

Expected

An alarm firing under in-memory storage should behave like the disk-backed
path (handler runs or a well-typed JS error surfaces), never abort the whole
process.

Notes

Happy to send the exact single-file config + worker used, or open a PR with a repro test if pointed at the right test layout.