Steve's TDR suggestion is the right starting point, and I would add why it works here.

PL-259 / SO-239 was designed for HF and never had an impedance spec - the nominal 50 Ω is really an average, not a controlled value. The interface geometry puts a fair amount of series inductance in the shield path plus a capacitive step at the mating plane. At 30 MHz that is invisible. By 400 MHz you can see the mismatch on a VNA.

A TDR makes this concrete because you get impedance vs. length, not just vs. frequency - you can see the inductive region at the connector and the capacitive step at the mating interface, and how long each one is. From there you can estimate what fraction of a wavelength each discontinuity becomes at your frequency of interest, which is really the question that matters.

Practical notes if you measure it:
- Keep the launch coax short and use the same connector type on both ports, so you are characterizing one interface rather than two in series.
- Set the reference plane right at the connector, not at the cable end - otherwise the cable loss and delay blur the connector response.
- Expect the ham-radio data you found to be all over the place; most of it was taken without controlled launches, which is probably the methodology issue you sensed.

Good question - and I think the test as described has a confound in it.

Two things change between a sharp 90° corner and a swept (radiused) inner conductor:
1) the local impedance discontinuity at the corner, and
2) the current path through the bend, so the conduction cross-section and current density differ.

The thermal camera only measures the result, not which mechanism caused it. A sharp corner crowds current at the inner edge of the turn, so localized I²R heating rises even if the impedance bump itself is small. A swept conductor spreads that path out, so it runs cooler for the same through-power.

If you want to separate the two effects, I would approach it this way:
- Run a TDR on both adapters with the same reference plane. That shows the magnitude and the length of the impedance discontinuity directly, rather than inferring it from temperature.
- Compare return loss across the band. If the swept part is genuinely better matched, you will see it above the frequency where the corner length becomes an appreciable fraction of a wavelength.
- Drive both at identical through-power and let them reach steady state. A single thermal snapshot is easy to misread if one part had better thermal contact to the housing.

So: not a wrong observation, but not a clean test either. Set up that way it mostly tells you the sharp corner runs hotter - which is expected - without proving whether that is mainly an impedance effect or mainly a current-crowding effect.