Functional components forget their watched atoms in strict mode

[matej-ibis-ai]

Running with React strict mode (enabled at the top-level component), React effects are cycled one extra time, which causes functional components created using reagent.impl.component.functional-render to dispose of their watched values, and not watch them anymore.

The effect defined in functional-render should provide a symmetrical action to ratom/dispose! in its first part, in the mount function.

In my preliminary tests, I have extended the Reaction type for the following two methods:

(_snapshot [_this]
  [state watching auto-run])
                                                      
(_restore [this [s w a]]
  (set! watching w)
  (set! state s)
  (set! auto-run a)
  (set! dirty? true)
  (doseq [watcher (set w)]
    (-add-watch watcher this handle-reaction-change)))

and used them to save the related values of the Reaction in a new Javascript field between the cleanup and second run of the effect, like this:

(react/useEffect
 (fn mount []
   (when-some [reagent-state (.-current state-ref)]
     (when-some [r-argv (gobj/get reagent-state "ratomSnapshot" nil)]
       (let [rat (gobj/get reagent-state "cljsRatom")]
         (._restore rat r-argv)
         (gobj/remove reagent-state "ratomSnapshot"))))
   (fn unmount []
     (when-some [rat ^ratom/Reaction (gobj/get reagent-state "cljsRatom")]
       (gobj/set reagent-state "ratomSnapshot" (._snapshot rat)))
     (some-> (gobj/get reagent-state "cljsRatom") ratom/dispose!)
     ))
 ;; Ignore props - only run effect once on mount and unmount
 ;; (which means always twice under the React strict mode).
 #js [])

That solved the issue I have described here, although I still see failures in expanding seqs rendered by these functional components -- but I will be filing that as a separate issue when I gain more understanding of what's going wrong there.

[matej-ibis-ai]

After another day of debugging learning Reagent, I understood two shortcomings of my tentative fix:

1. In the _restore method, the dirty flag should be set to false; setting it to true prevents this Reaction (sitting at the core of the functional component) from triggering rerenders, which is the opposite of what I want when restoring the thing.
2. The dispose! method is still not perfectly undone by _restor[ing] an earlier _snapshot of the Reaction object because the -remove-watch implementation calls dispose! recursively on any watched items if it was their last watch(er).

The first one should be easy to correct -- simply by setting dirty? to false instead.

I will have to think more about correcting the second one.