Filled (nonconductive) epoxy.
What's bad about rigidity? Thermistors are ceramic chips internally, nothing bad about it. Long as you have strain relief on the leads of course.
Performance is entirely described by the thermal time constant, and what error if any arises from dissipation off the outer surface of the sensor (or blob of goo it's inside), and how that stacks up against the thermal resistance of the "Blob-O-Goo" (as a professor I had would say), contact resistance, and wherever the sensor is in the midst of those temperature drops.
Heatsink goop is a good idea for mounted types (screw down / etc.), though not really necessary, the heat isn't going anywhere. It'll probably improve the time constant and error a little, but I don't think I would bother if it were a production kind of thing. If it's that precise, use an embedded thermocouple; you'll get better guaranteed precision and accuracy.
I wouldn't suggest rubbery goo for the most part, since the conductivity isn't going to be as great. That said, as long as the sensor is at the bottom, close to the thing-to-measure, it needn't be too bad, and the insulation could even help reduce error.
For spot testing with thermistors, this is what I do:
- 0603 chip thermistor (1% 10k, R85/25 around 3500 -- a pretty average, highish tolerance, NTC thermistor)
- 3" pigtails: 37AWG enameled wire, tack soldered to the chip
- Hookup wire leads back to the DMM, resistance range; use lookup table to find temperature. Or, set up your own temp measuring gizmo with an ADC (8 bits is marginal, 10 bits good enough, 12 bits plenty) and the lookup done in software.
Use: dip the thermistor in heatsink goo, then touch it to the DUT.
Time constant is quite rapid, and temperature pretty faithful. It does "load down" small objects (like SMT transistors, which have about as much heat capacity as the thermistor chip), but works just fine on big things (heatsinks, etc.). Makes a mess (bops of heatsink goo everywhere). Beware of conductive objects shorting out the thermistor (maybe hold it lengthwise to the surface rather than broadside, so it doesn't get shorted out), or being shorted by it (don't use near device pins, etc.). Not so great for very high temperatures (>100C), since the resistance is already very small by then. If you have extended temp data, it'll be good out to 180C or so when the solder begins to melt (heatsink goo is quite robust, in and of itself, by the way). Obviously, more than that and you have little choice but to use a thermocouple.
Tim