What is a best practice for linking Just-In-Time Tasks?

Sure, date constraints, especially hard ones, are not recommended. All predecessor link types except FS0 are not recommended. Lag, especially negative lag, is not recommended. They are all "constraints", some looser than others. The reason these are not recommended is because the use of them, all mixed together, narrows you options, and of course usually results in a great big tangled mess. However, they all have uses in particular special cases. Even manually scheduled tasks, definitely not recommended, have an occasional use. If you use the SF you will get what you want to display as your "intention". For a bit of extra spice, give it some negative lag as well, so that prep task can't finish until 3 days before important task starts. Can even go SS with negative lag.

But displaying the intention is not all there is to it.  That has to follow from the basic question that the network modelling is supposed to answer first, which is how soon each task can start and, secondary question, how much float does it have. Before you make that SF link you have that answer, but that information disappears when you make the SF link.

There is another way. You already have the information you need without the SF link. The usual Gantt chart view has the early bars going from the start (that's the earliest start) to the (earliest) finish. But you can make a copy of that view, call the copy "AA Late Bars View" and reformat the bar styles in that view to go from the latest start to the latest finish.

Yes, absolutely.  Years ago I did a couple of training classes for two groups of project managers at NASA.  When we talked about the SF dependency, they described how the launch of a rocket is a "sacred event", whose date and time are "etchec in store" so to speak.  They have a final milestone in their projects that represents the date and time of the launch.  Then the final sequence of activities that must be performed before the launch are all linked from the launch milestone backwards using the SF dependency.  So, your use of SF is reasonable, given the situation you describe.  As an aside, when I finished my classes with the NASA project managers, I asked them how I could volunteer for the first manned mission to Mars.  And they all laughed at me!  This was LONG before Elon Musk and SpaceX, by the way.  HA!

I understand your hesitation in not wanting to use a Start-to-Finish (SF) dependency, because it means you are scheduling backwards.  In fact, several of my clients PROHIBIT using an SF dependency.  However, I am a Baby Boomer, which means I believe "Rules are meant to be broken."  In your situation, the SF dependency is the only dependency that actually models the scheduling behavior your are seeking.  Therefore, I would recommend you use the SF dependency in your situation.  Hope this helps.