On the other hand, bringing the primary and secondary to the same potential increases the inter-winding parasitic capacitance - but then makes the test set-up much more resistant to stray capacitance. My friend who had been doing the capacitance measurements was using floating (actually battery-powered) meters and oscilloscopes to do his measurements, and was pleased with the resulting good high frequency response. But is this a realistic test set-up?
Any amplifier that has feedback connected to the secondary requires that the secondary be at or very near ground potential. Amps without feedback can have their secondaries floating - but this is a safety hazard. A common failure in transformers, especially ones with lots of interleaving, is a primary to secondary short. With a floating secondary, the transformer will still work, but the speaker terminals will be several hundred volts above ground! Thus I have always recommended that one side of the secondary be grounded, even in zero-feedback amplifiers. If a short occurs, then (hopefully) a fuse will blow. Another problem with a floating secondary is that it puts high frequency common-mode audio signals on the speaker leads. Stray coupling or capacitance to metal and wiring in the listening room will alter the high frequency behavior of the amp.

Recommendations
After reviewing the issues above, it is clear that for consistency in measurements and conformity to most amp designs, the primary and secondary need to be tied together for testing, as shown by the dashed lines in figure 1. Using a source resistance of 1/2 the rated load will simulate the situation in a real amp the best, but some people recommend using the same source resistance as the rated load. This is more conservative, and will degrade the low frequency response by about 25% and the high frequency response by 5 to 20%. To avoid ground loops, one side of the test set-up should be floating. If there is a choice, it should be the side with the least capacitance to

    In order to see the effect of no interleaving, an anonymous push-pull transformer that happened to have 3K from the center-tap to one side was dug up from the junk box. (For reference, it is stamped "320085-1-X".) The secondary is wound on top of the primary with no interleaving at all. Figure 10 shows the response - but note the different scale. The measured primary-to-secondary capacitance is only 210 pF. There was no difference at all between the A-P input being grounded or floating and only a slight difference when the primary and secondary was connected together.

The Culprit
    The degradation in high frequency response when the primary and secondary are tied together (the dashed connection in figure 1) is the worst in the transformer with the highest primary-to-secondary capacitance and almost non-existent in the transformer with low capacitance. When the secondary is allowed to float, this inter-winding capacitance brings the secondary to an intermediate AC level, and lowers the effective parasitic capacitance. When the primary-to-secondary capacitance is large, variations in the stray capacitance to the generator ground can have a large effect, hence the variations in response with different meters. Unless a special test fixture is made that presents very low and standardized stray capacitance, it looks like the measurement consistency with a floating secondary is hopeless.