So a lot of things to unpick here:
- The kit from SDR kits is really just the normal Rosenberger parts with a box and a 4-wire resistance measurement. I seem to recall that the price on Mouser was around the same, but if you can get them cheaper elsewhere now, no reason to pay more.
- The offset lengths given by HHFT and Kurt Poulsen contain the first capacitance/inductance term (C0/L0), which really just gives a delay. You cannot derive them from a mechanical measurement. You also cannot measure them on a VNA without calibrating it first against known standards.
- I have seen some measurements on a lab-grade VNA of the open and short and they didn't look bad to 5GHz+. The measurements shown on the HHFT page are probably not a bad indication of what to expect. On the plots they show, the performance >3GHz seems to be limited mostly by the load.
- The mechanical dimensions of the parts are probably quite precise, but the properties of the PTFE dielectric will vary from batch to batch. So if you need a more accurate characterization (with higher-order capacitance terms), you would really need to measure each one individually (like Kirkby does). Or use 3.5mm parts with air-dielectric (like Keysight, Spinner, Maury Microwave or the proper Rosenberger cal-kits). Those are very consistent and the RF properties can essentially be calculated from the CAD model.
- I would recommend against putting a male cap on the female thru used for open calibration. The fringing capacitance is already factored into the offset delay and is likely to increase (slightly) with a cap. The point of shielding the open is mainly to prevent radiative loss, but even on a male SMA (which essentially has a tiny antenna if left open), this isn't really an issue until at least 5-6GHz.
- The load has a relatively large tolerance (all of those cheap general purpose loads have). It is only guaranteed to have >26.4dB return loss above 2GHz. In reality, most will be a lot better than that but don't rely on it. At low frequencies, the main contribution is the tolerance of the resistive film, so selecting by DC resistance is not a bad start if you have several. Entering the DC resistance as a parameter is usually also better than not (at least to 2-3GHz), but keep in mind there is also a short piece of transmission line between the connector and the resistive film, so changes in the resistance don't translate into a purely real impedance change at higher frequencies. If you buy separately, any 12-18GHz female load from Minicircuits, Huber+Suhner, Telegärtner etc will be very similar, in case you can find some of those cheaper.
- Buying a pack of cheap random loads on eBay and selecting those by DC resistance may end in disappointment. Some of those have really poor RF properties and even if you find one with 50.00Ohms at DC, it may be way off at a couple hundred MHz, let alone a couple of GHz. Buying cheap RF parts of unknown providence is risky business in general. Mechanical tolerances may be off and it's easy to damage a good connector by mating it with a bad one. So inspect carefully, keep questionable parts separately and don't connect them to expensive instruments/parts. Use some sacrificial adapters or cables in between. Never use male SMA connectors where you cannot rotate the barrel separately from the center pin (like in some cheap "calkits"). That is a sure way to eventually ruin the mating female connector.
Overall, if you are looking for something cheap for non-critical measurements, the Rosenberger kit is probably your best option. It will not give you the confidence and accuracy of a proper calkit, but you get quality parts (mechanically) and to 2-3GHz performance will likely be respectable. It is also a female kit, which is likely the more useful one if you are getting just one. I don't know why all the cheap NanoVNAs all come with male ones.
I am a bit more sceptical about their male kit. Like I said above, the male open cap doesn't really do much at low frequencies anyway, so the open measurement will depend more on your test connector than on the calkit.