I have the beta version of join

In the lab.  Production release with the following changes:

Code deleted and table of devices gone, full inheritance of cursor enable. Less code in my version.  Devices needed initialize their own cursors, so I changed the device dll entry point to a single entry for set up and configure of cursors.

I added the new memory device INDEX, because the html scrapper used it. The code was available, I made the code one of the two types of structured memory.  Then the scraper just inherits that type of device and modifies the step and skips.

Devices have a mode identifier to specify if the key is fixed size or variable size. .  Devices can agree teat the key variable differently, join loop remains unchanged.

The shell is now full duplex, delivering an array of linux style arguments and accepting array of the same in return.
The call is"

ExecMyDevice(int * in, int * out, char *  argv[]);

Any subsystem receiving a call, via command can:

int argi=*in;
char * str= argv[argi]; argi++;
int i = atoi(argv[argi]);argi++;
*in = argi;

On return output the opposite:

int arg *out;
out--
argv[argo]=malloc();

*out=argo;

The html deice has become simpler and is a better development model.  Simple create the standard memory methods and intercept the call backs, apply device specific matching, then use the underlying memory methods to direct navigation. In this manner, searching html is the same as searching Websters standard dictionary; both adapted to LazyJ search syntax.

Code eliminated, demoted. Device development simplified.  I have a production version, unreleased.  The Entire staff here at Redneck Inc was hit with tax an plumbing issues, at the same time; and we are undergoing a system upgrade.