Notification test cases, for the opt out that only half worked

A notification preference is a promise across every channel, and it is usually implemented per channel by different code paths written at different times. That is why the characteristic failure is partial: email respects the setting, push does not, and the user experiences the product as ignoring them. Testing it means triggering the notification and checking all four surfaces rather than the one the feature was built for. Per channel granularity matters too, because a single global toggle forces people to choose between every channel and none, and most of them choose none.

The classification boundary is where this becomes a compliance question rather than a usability one. Transactional messages must be delivered regardless of a marketing opt out, and marketing must not be sent as transactional to get around it. Both directions fail in practice. Sending a promotional offer dressed as a service notice generates complaints and eventually blocklisting. Suppressing a genuine security alert or a billing failure because of a marketing preference is the mirror image, and it is worse, because the user misses something that affects their money or their account.

Unsubscribe deserves particular care because the link travels unauthenticated in an inbox. It has to work without sign in, and it must be scoped so it can only unsubscribe, never reveal or modify account data, since anyone who obtains the email obtains the link. Its scope needs to be documented and honoured: unsubscribing from one campaign at a time until the user gives up is technically a working feature and practically a complaint. And the opt out has to survive imports and syncs, which is where it most often gets quietly reset.

Finally, the delivery layer runs on at-least-once queues, so duplicate sends are the default rather than the exception, and deduplication on an event key is what prevents them. Batching stops fifty events becoming fifty messages. Digests need to fire at the intended local time for each recipient, which means handling a daylight saving transition without producing two in one day or none. And a test send from a non production environment must be structurally incapable of reaching the real list, because that is among the most damaging self inflicted incidents available.