Gotcha, if I understand correctly, the recommendation is not to try to do interaction testing. I think this is actually pretty reasonable given that Blockly itself is well tested.
One thing I'm currently attempting in my tests right now to get more of an end to end experience is the following:
1. Create a blocky XML with my custom blocks
2. Call my code generator to generate code from the custom blocks (Lua in our case)
3. Run the generated code through the Lua interpreter with some stubbed methods that output the calls
4. Validate that the output was what we expected.
I originally spent some time trying to validate generator output with snapshots, but small changes would result in all the snapshot tests breaking, so I'm hoping that this scheme will prove a little more robust.
An added benefit to this method laid out above is that it should be pretty reasonable for more long-term regression testing, in addition to making sure that generated code results in a syntactically correct program :D.
Best,
Eric