This is in a way "good news", because I noticed the same trend in using spherical vs PBC/Ewald summation. When it comes to NMR exhange correlation (order parameters), it turned out to yield rather similar patterns/residue on the protein in (TIP3P/SPC or SPCE/E). But as you stated, smaller systems are probably more sensitive to such issues.
Since we're now in a wealth of computer power, I feel it is worth the extra CPU-cost to get answers closer to reality..
Please keep me informed about the TIP4P progress.
Thanks for the ref btw.