Make that choice explicit first. I've seen a useful teaching substitute become misleading after the real environment changed and nobody knew which set of responses the lesson still described

And record where the response examples come from. A saved value, a documented example and a deliberately malformed input can all be useful tests, but they establish different things about the target interface

Maintainer identified the supported version. I'm building the small interface we actually consume, with documented responses and separately labelled malformed cases. Unknown methods will fail.

Include valid negative status, not only success and broken reads. Those need different screen meanings.

What about a response that never arrives? My worry is a screen that handles every returned value beautifully but freezes while waiting for one. Can your software tests exercise the waiting and timeout behaviour without using the actual controller?

That's a different test boundary from checking the decoder, Ben. Still worth doing, but don't make Toby build a whole pretend network just to return a failure payload.

You can control completion in the test harness without pretending it is a network. I'd keep the decoder examples and the delayed-operation tests separate.

Who reads the timeout state during a shift? I'd want the maintainer to see the screen examples, not only the test names.

Strict response tests found an argument mismatch. Fixed that. The delay test is separate as Dinesh Abbott suggests; it exposed a wait with no display-age update, so we're not done.

Felix, fair enough; I meant control the response timing, not emulate cables and switches. Toby has found the particular screen behaviour I was concerned about.

Does the test also deliver the held response after the application has already timed out?

For that case, inspect which request the late response belongs to. A late answer can be valid evidence about an older request without being a current fixture reading for whichever request is now displayed

The held old response replaces the newer display value. Reproduced with explicit delivery order. I've left that failing test in place while the maintainer reviews request identity through our adapter.

Useful to know why it fails, but don't let the strict mock's passing cases become the headline in the handover. The person trying this version still needs to know the display can accept an old answer

Keep it offline until that is settled. Controller access returning won't repair the application logic.

Offline only. Maintainer found request identity discarded in the adapter and has a repair under review. Timeout wording and sample age are corrected; late-response regression still fails on the current version.

Thanks for the distinction. Has anyone checked another restart during that delayed-response test, or is that still outside this set?

Outside the completed set. Added it to the review list, not to the passes. The substitute now catches real software errors; it hasn't finished fixing them.