Unlike [JENKINS-14514] this is a true fix rather than a workaround (now removed), and is more general.
cjo9900 discovered that behaviors were being redundantly registered (as of 1.474 the monolithic JS is broken up);
this caused some behaviors to be run repeatedly on the same elements, breaking reasonable expectations of some behaviors.
The ideal fix would be to change Behavior.register to be idempotent: for example, key it by selector, then maintain a set of distinct behavior functions for each.
Unfortunately some adjuncts directly call Behavior.list.unshift, bypassing register(...), which would be tricky to intercept (would need to make a mock of Array).
The known one cases are in core, but it is possible plugin adjuncts do this too, in which case it would be incompatible to (say) change the Array<Map<String,Behavior>> to a Map<String,Array<Behavior>>.
Instead, permitting redundant registrations as before, and just silently skipping all but the first at runtime when applying behaviors.
Beware that since adjuncts are loaded from multiple places, different JS function objects are registered each time, so a naive set of behavior functions does not work;
have to identify functions by their toString in order to ensure that each is run only once.
(Currently once _per selector_, conceivably >1x per element; could if necessary be refined to make sure a given behavior is only run once on a given element during one call to applySubtree even if the element matches multiple selectors.)
One of the problems with the current way Jenkins puts JavaScripts is that most everything is in a single file. It doesn't highlight the relationship between those fragments and their corresponding tag files.
So I'm experimenting with moving out these fragments into individual JavaScript files co-located with the tag file. They then get loaded into the page via <st:adjunct>. Each script is served with unique URL such that the browser will not even have to try conditional GET when loading different pages, so this should prevent the performance problem caused by the fact that there are now lots of <script> tags in the page.
If for some reason even this turns out to be a problem, we can always add additional build steps to concatenate them all.
This makes it consistent with other buttons like repeatable-add-button, and use of '-' prevents name collisions with form names, and make all of those UI controls addressible in the same namespace
Previously, the code was "find the nearby element of the expected @name,
then make sure it has gotten the right parent."
So if the structure is like:
{/* outer object */
name: "parent-name", /* A */
child1: {/* inner object */
name: "child-name" /* B */
},
child2: {/* inner object */
name: "child-name" /* C */
}
}
... and if we are looking for "../name", it finds B, and it gets
rejected, so we won't be able to find A, which is what we are supposed
to find.
With this change, we keep trying until we find the right one,
so the above case will correctly return A.
The actual release needs to happen in a lock step, as the new plugin
needs to require yet-to-be-released version of the core. So for the time
being it's a snapshot dependency, deployed in the snapshot repository.