For 100% procedural generation plants, making/defining a small set of parameters to apply for the biome type, based on statistical distributions of realworld plant morphologies could work-- (EG, "70% of realworld desert plants are of X morphology, 10% of Y, 5% of Z, 5% of A, 2% of B, and 3% of C"-- then have the computer pick a classification for basic morphology type based on target biome, then based on this initial decision, pick some other features, again based on statistical spreads-- EG, Some percentage of type Y morphology has Q type flowers, which would define # of seeds produced, etc.)
Applying that to the "original plants" would make a good trial and testcase for the procedural generator's quality.
Total scrapping of those crops just seems kinda wasteful; they could be useful as testbeds for a procedural generator.
This can be taken a step further, where different characteristics are correlated with each other. Producing correlated random numbers involves Cholesky decompositions and other arcane math tricks, but Toady actually speaks math so that's not the barrier. The barrier would be coming up with a reasonable matrix (most likely a community effort like the materials properties mentioned above).
With procedurally generated creatures, having no idea what to expect is part of the fun. But with procedurally generated plants (with which in-game civilizations have centuries or millennia of experience), will there be a way for the player to get an idea of what a plant is, its farming requirements, products, etc.?It would be a shame for Urist McFarmer to spend a couple seasons tending Moonstone Bushes to feed the fortress, only to find out there are no edible parts.
Edit: grammar