Fixed incorrect de-allocation of a classloader from the exported object
table.

The fix is a defense-in-depth; it prevents classloaders referenced in
the object graph from doubly released, then we also make it impossible
for bugs like this to deallocate the key classloader.

Some of our tags, such as <optionalBlock> and <dropdownList> involves grouping of sibling table rows, and controlling visibility of them. So when such tags nest to each other, the visibility updates need to be done carefully, or else the visibility could get out of sync with the model (imagine outer group is made visible while inner group is not visible --- if all the rows are simply enumerated and visibility changed, we end up making the inner group visible.)

The rowVisibilityGroup object in hudson-behavior.js is used to coordinate this activity, and this test ensures that it's working.

Added autocompletion to build after other textbox, and textboxes can now use autocompletion without needing a true field via the autoCompleteField attribute.

on IE8, <A> element always gets empty string as name, not null. Empty name is not valid any way, so detect that accordingly.

in IE, innerHTML of TABLE is not modifiable (See http://support.microsoft.com/kb/239832), so this doesn't work. Looking at the similar code in the latest prototype.js, a work around is to create the full TABLE HTML fragment and inserting it to DIV.

in Firefox 3.6 on amd64 Ubuntu that I tested, setting innerHTML of TBODY to "<tr>...</tr>" doesn't just add a TR, and Firefox tries to fix it to something broken. But setting innerHTML of a TABLE to "<tr>...</tr>" works as expected.

that results in missing tail bits when the master is heavily loaded.

strange intermediate nodes.
(For example, when we are inserting a number of <TR>s)

making Behaviour.applySubtree to consider the given node itself as a possible match. Previously it wasn't.