I made a mistake, I did manage to find some +/-0.05 degC thermistors for USD4 on Digikey, so the USD9 figure can be feasible in low quantitites. I actually bought a bunch of the thermistors you linked to; the part variation is quite high so calibration is needed to get good results. I think this is one area where we can truly say Cheap, Accurate, Simple: pick 2. Edward Mallon whom the authors also cite https://thecavepearlproject.org/ has blogged extensively on this.
Personally I would try to avoid analoque devices because I don't want to deal with having to amplify it, which is where the errors creep in. Maxim make an ADC that directly compares resistance ratios, and HX711 simplifies bridge measurements. I'm not sure how this logger does it; the Arduino's ADC isn't very good as you said. The MAX30205 human body temperature sensor may be worth looking in to for this application. The measurement range is small but probably adequate for this application, and the accuracy is well documented.
The sensor time constant is a bit high but should be ok for for this low-rate application, and kudos to the authors for characterising it. There are applications where the time constant needs to be <1s, like for ocean turbulence structure measurement, but that is more physical oceanography and only indirectly wildlife conservation. The internal sensor simplifies the logger and this is a big advantage in this instance, one I have taken advantage of in the past.
I have considered PET bottles in the past and it is very attractive. They do very well with internal pressure (7 bar I think!) but with external pressure they are prone to buckling. One approach may be to re-mould the bottle to be smaller so the walls are thicker. This should be easy because the bottle shrinks with the application of heat anyway.
I like cheap hardware too! Plus it goes well with citizen science in a field where investigation is labour intensive and it lends itself well to outreach efforts.