I decided to use the more direct method, of disconnecting feedback to the EX laser PZT, and then looking at the cavity flashes. 

Attachment #1 shows the cavity swinging through two resonances (data collected via oscilloscope). Traces are for the demodulated PDH error signal (top) and the direct photodiode signal (bottom). The traces don't look very clean - I wonder if some saturation / slew rate effects are at play, because we are operating the PD in the 30 dB setting, where the bandwidth of the PD is spec-ed as 260 kHz, whereas the dominant frequency component of the light on the PD is 430 kHz.

The asymmetric horns corresponding to the sideband resonances were also puzzling. Doing the modeling, Attachment #2, I think this is due to the fact that the demodulation phase is poorly set. The PDH modulation frequency is only ~5x the cavity linewidth, so both the real and imaginary parts of the cavity reflectivity contribute to the error signal. If this calculation is correct, we can benefit (i.e. get a larger PDH discriminant) by changing the demod phase by 60 degrees. However, for 230 kHz, it is impractical to do this by just increasing cable length between the function generator and mixer.

Anyway, assuming that we are at the phi=30 degree situation (since the measurement shows all 3 horns going through roughly the same voltage swing), the PDH discriminant is ~40 uV/Hz. In lock, I estimate that there is ~60 uW of light incident on the PDH reflection photodiode. Using the PD response of 0.2 A/W, transimpedance of 47.5 kohm, and mixer conversion loss of 6dB, the shot-noise limited sensitivity is 0.5 mHz/rtHz. The photodiode dark noise contribution is a little lower - estimated to be 0.2 mHz/rtHz. The loop does not have enough gain to reach these levels.