In Hauptwerk v2, all internal processing will happen at 32-bit, and output will be at the highest resolution that the card supports (e.g. 24-bit), regardless of the resolution of the original samples. Hence the overall resolution could be much higher than that of the samples, without requiring any increase in memory. You would only need to increase the memory requirement if using true 24/32-bit samples, or higher sample rates.
Am I correct in understanding that even using 16-bit samples, it will be awhile before PC's will allow enough polyphony to use wind pressure modeling with stereo ambient sampling of larger organs? In other words, does this question boil down to priorities of what we want to do with available CPU performance? Do we prefer to stick with smaller organs and/or dry mono-samples in order to have higher S/N using 24/32-bit samples
Use of 24/32-bit samples doesn't really add any processing overhead; you just need more memory. Stereo adds some overhead but nothing like double. Ambient recordings (long reverb tails) use up more of the available polyphony, since they have to continue playing after the key is released. Wind pressure modelling also adds processing overheads, and thus reduces polyphony, but, provided that my estimates are correct, should still be feasible with 1000+ polyphony and stereo ambient samples (at any resolution).
A polyphony of 1000 is always my target - I think that is adequate for pretty much any realistic scenario. My aim is to use all available processing power on the latest processors to produce the best sound I can, whilst still achieving that target.
One other point - for small sample sets (e.g. positive organs), there is no real reason not to use 24/96, since the memory requirements will be small anyway.
Martin.