You'd want (at minimum) two bytes per tile, I think:
Byte #1: eight bits, each specifying whether a given tile (up-left, up-center, up-right, middle-center, and so on) is part of the same structure or not.
Byte #2: type of structure. Tubelectric, for instance, has multiple wall types; you wouldn't want to create a wall of mixed purple and red tiles by accident.

Although, I wonder if it might not be better to separate such things (i.e. your idea and mine both) from the .j2t files, for two reasons:

• Levels may potentially need their word data intact for the gameplay to work properly. If you get a level in your cache, ideally you should not be able to break it just by saving it in MLLE/WebJCS. Info about how tiles are related to one another, however, is by no means necessary.
• Adding such information to old tilesets, including the official ones, would mean a whole lot of recompilation. This is bothersome to do, leads to version control issues, and may in some cases not be permitted by the tilesets' original authors. Setting up separate files (.j2tdata? whatever) means smaller filesizes, no direct modification of the tilesets themselves, and moreover it's easier for multiple people to refine the info of a given tileset, either for general release or for their own purposes.