Well, MOSFETs aren't very much better sorted (BS250 Vgs(th) -1 to -2.4V). Again, physics... to a certain extent. FETs are FETs, though within a broader meaning this time (Vgs(th) varies over a relatively wide range for both kinds, though for different reasons).
Using JFETs here would amount to removing the positive gate bias resistor, and probably adjusting the source resistor. Yes, variable or selected resistance would help, since Vgs(th) still has a wider range on most types.
For low noise applications, matching isn't really necessary, for the properly spec'd types (e.g., 8-20mA -- not the pitiful 2-20mA range you get on floor sweepings like the 2N3819!), and your results may be even better than expected if they happened to come from the same lot, or wafer. (The noise reduction goes by statistics on Gm, which is about proportional to Idss. If you run them at Idss, the ~2:1 Gm spread is only a sqrt(2) spread in statistical weight, still enough to care about, at least until you have, like, 6 or 8 in parallel.)
If you're going to kind of give up and use an op-amp, the excess noise is probably about what you'd get from a discrete MOSFET, so not a big deal either way -- if you have the op-amps on hand, who cares?
That said, there are some quite good JFET op-amps available (if you don't mind shopping for them), which will surely beat discrete MOS, and probably even a crummy JFET stage too. Which is pretty cool, considering op-amps have on the order of 100 active transistors in them, these days.
Changing just the first stage is nice, because that has the most impact on noise floor while making the least on distortion response, except for the very strongest of input signals.
Tim