I want to give real-world advice as opposed to some of the more theoretical suggestions (the mention iPads doesn't quite line up with "small cheap computer").
A lot depends on unspecified contextual factors, so I'll just throw a couple ideas at you.
If you have a cache of sub-GHz hardware already lying around, I would recommend Win98 and QuickBASIC 4.5. The ability to hit F5 and immediately get output is excellent for reinforcement learning, and being able to CTRL+Break out of unruly mistakes gone rogue great as well.
After a few months (or only weeks?) of QuickBASIC I would move up to Visual Basic.
To be entirely honest, I stumbled on VB 1.0 while browsing at a library many years ago, threw added it to my CD to take home "for the lulz"... and ended up writing a ton of things in it. I found 2.x-5.x's UI to be unnecessarily full of things I didn't want to know about at the time. It _is_ very Spartan; this may appeal for a very short time.
After the novelty of drawing up terrible UIs has worn off, throw them a curveball in the form of VB 6. They'll have the fascination to chow through the additional UI complexity, and VB 6 occupies the sweet spot of being the last in the pre-.NET series and a good preparation for Visual Studio...
...which will need to be introduced at some point, presumably along with an OS switcharoo from Win98 to Win10. My recommendations for such archaic platforms are not without due consideration; the "total mental model surface area" of Win98+DOS is orders of magnitude smaller as Win10, which may provide a friendlier learning environment than "hi you can play in the rocket capsule but don't do anything that affects the engines underneath or you'll blow us all up and I'll be really mad." The very opposite will be true since it's a "toy" environment, so the kids would be free to kill everything as many times as they wanted/needed (within the limits of your patience ;)), and this will fuel the inspiration to discipline their interest in discovery and learning into something usable.
A microcontroller-based environment does provide some of the kinds of bare-metal experience that using QB on an old PC does, but only if you completely ignore tooling and infrastructure. Unfortunately I have insufficient (...any...) experience so can only really comment on the difficulty of this aspect.
The above reflects (part of) my own history (I unfortunately skipped the bit about VB6, which was why I had so many things in VB1.0. I only ended up learning how I'd even go about finding a copy long after the window it would have been useful had passed, so perhaps I'm a bit biased about it.)
The one thing I'm trying to consider, though, is the thing about total surface area, for want of a better phrase - I can remember struggling to figure out JavaScript using IE 5.5 because I didn't know anything else even existed at the time (in 2003, no internet) and perceiving the existence of and hating the browser sandbox. The devtools REPL makes things so different nowadays but it's undeniable that a browser is a gigantic pseudo-operating system on top of another operating system and most problematically this is viscerally perceptible; it's not seamless at all. There's no sense of true freedom, that you're pushing what you're controlling to its limits. (Incidentally speaking, this has some interesting parallels with the Web's recent transformation into the media format of the day, its adoption of DRM, etc.)
The same is true to some extent with QuickBASIC, and this is why I would recommend using it for a maximum of 6 months, but only allowing someone to use it for that long if they had learning difficulties. I personally believe that my own development was severely impaired by the fact that I had nothing else available when I was younger. I'm only just breaking even in terms of having a mind able to make use of CS first principles, but in all fairness this is predominantly due to my autism and not entirely QuickBASIC's fault ;)
A short, engaging path straight to assembly language might be a good idea. I'm still yet to wrap my head around this particular subject, nearly 20 years on, because I'm still un-learning the very complacent computer science models I unfortunately ingrained when I was very young.