some layers can feed information forward, and some back.
for example, if I'd like to train a span classifier to identify railway station, it stands to reason that many of the features captured by NER can be very useful features feeding that model.
conversely, if I train (no pun intended.. ) a entity linker to identify the relationship that includes a train station (after it had been identified) then that model can feedback into the previous layer, and correct that.
(or document Classification feeding back to earlier layers)
with much of the explicit process being abstracted away (and thank you for that) by prodigy recipes, is there some reference material that might help me avoid pitfalls and gain advantages where such feedbacks may affect the model.
(or is it the SOP to run forward, and use a model to create a new dataset, and then train preceding components on that for example. )
As of spaCy 3.1 it is possible to propagate predictions (Training Pipelines & Models ยท spaCy Usage Documentation) forwards (not backwards) between components however, not all predictions can be used by all models. For example, the spaCy textcat architectures do not use NER predictions. You can always control for it by e.g. implementing your own model and using it as a subclass of spaCy trainable component. Although it doesn't deal exactly with the problem described, this tutorial shows how to implement a custom component and a custom Thinc model and might be useful. Other than that spaCy discussion board is definitely a good resource on these topics.
In general, for more complex training protocols it is recommended to use spaCy directly as it removes some of those abstractions you mention and gives you more direct control over training configuration/debugging. Prodigy train command is a thin wrapper over spaCy train and the default components are mostly independent, precisely, so that the users don't have to worry about the pitfalls you mention.