After an intense weekend of debugging, I've discovered that the Lucent driver actually (sort-of) DOES work!
The scenario: all cards and OSes have the 21-July release of Lucent's drivers and firmware. The wvlan_cs driver loads up, starts the net device, but can't talk to windows boxes. The wavelan2_cs driver won't even load up.
It turns out that the reason the wavelan2_cs driver doesn't load is because cardmgr dies BEFORE it can send the device a DS_BIND_DEVICE ioctl! Stepping through cardmgr with a debugger, it shows all the signs of classic memory corruption. Harmless calls like "syslog" lock up forever, and previously valid instructions cause "illegal instruction" and "segmentation fault" errors.
So... when using the wavelan2_cs driver: if I jump over all the "syslog" calls, which should theoretically not make any difference, the net device is correctly brought up by the module. Yay!
I still can't talk to the windows box, but there may still be something funny going on. Even thought I delicately stepped through cardmgr with my debugger (and things inside it were corrupted), it reported several "could not adjust resource" (memory and I/O errors).
So, until I (or David!) can track down the corruption in cardmgr, it may be possible to use Lucent's driver with the latest firmware to talk to windows boxes...