I really should have broken it up into several monotonic commits, and in fact I would have done so if I'd known how I was going to solve the problem when I started.

Here's a summary of the conceptual things I changed:

• Changed some BiConsumer<BlockCreator, Expr> to BiConsumer<BlockCreator, ? super LocalVar>
• Added a caughtName to name the variable of the caught exception
• Made sure that division and remainder operations are always "bound" even if their inputs are not, because they can throw an exception (we could narrow this down to only div/rem on integer types) - the basic defect here is that an unused div/rem will get deleted otherwise, losing the side-effect of checking for division by zero (not that I recommend doing n/x as a test to see if x is zero, but correct is correct and incorrect is incorrect!)
• Extracted all anonymous subclasses of Goto to top-level classes
• Deleted some unused code (incl. the incomplete/nonfunctional Cleanup idea)
• Modify Goto so that the target label method accepts the "from" block to support finally
• Add new CleanupKey concept which is the primary interface for instructions which have to run cleanup on the way out of finally blocks (this includes the Goto subclasses plus Return, but not Throw (or method calls or other ops which may throw), which is always handled by a catch-all block)
• Separated try behaviors into two separate Item subtypes: one for try/catch and one for try/finally; TryCreatorImpl is no longer an Item and instead composes a combination of one or both of these subtypes after try_ completes
• Update TestClassMaker so that ClassFormatError also triggers a nice class dump
• Testing

I spotted and understood the first few described changes, but the gist of the issue, not really. I don't even know what "redo" is :-) Maybe you could walk us through this on our Monday call?

Sure, that works for me.