I put my HC-SR04 sensor in the mail to Tim today. Hopefully it'll tighten the test cycle and benefit others who have HC-SR04 sensors. 
My backup HC-SR04 is definitely less reliable than the one I shipped out, but it still "works" (for some value of "work"). I have digital filters in place to try to determine if a reading is true or not, so this mitigates the inherent issues with the sensor.
My code base can also create a virtual sensor from N actual sensors, and I was planning to do this to work around the unreliability, since it seemed unlikely that two sensors would be unreliable at the same moment in time (unless there are common environmental or power factors that cause this). If the HY-SRF05 is more reliable, then I can probably start trusting the data more and tune the digital filters to be more lenient.
Is there any demand for virtual sensors as I described? If so, I can work on turning what I have into a subclass of NewPing. It sounded like other people are using multiple sensors in non-redundant ways, and if the reliability solution really is "just buy an HR-SRF05" then maybe that effort would be wasted.