BTW I also think it very likely that whenever there are (avoidable) chordal losses, (through using a setup with small chainrings and sprockets to give a certain gear ratio), there are likely to be other losses that are not accounted for in most transmission efficiency tests.
Specifically I am referring to the fact that when there is any chordal effect, either the output or the input has to vary in speed as each roller passes over each tooth on the sprocket. My bet is that, on a real bicycle going at a brisk pace, it is not the output that varies in speed, but the input.
If so, one's pedals, cranks, feet and legs are fluctuating in speed which is unlikely to be a process that conserves energy (in much the same way as riding on a bumpy road doesn't either). The smaller the sprocket the greater the variations in speed that are seen.
For example if riding 36/12, (assuming each power stroke occupies a quarter turn of the cranks) there will be 36/4 = 9 fluctuations in speed during the power stroke.
Each fluctuation's size (in percent) is given approximately by
100x (1- Cos (360/2n)) where n is the tooth count on the sprocket.
In the case of a 12T sprocket, it is about 3.4%.
If the same gear ratio is achieved by using 45/15 instead there are ~10 fluctuations of ~2.2% instead, and if the gear is 55/19 (which is not the exact same ratio, but close) there are ~13 fluctuations each of ~1.35%.
Because the change in kinetic energy of the fluctuating element is dependent on changes in (V1^2 - V2^2) this effect is unlikely to be insubstantial. For example if V2 is 3% less than V1, the change in kinetic energy is of the order of 6%.
So if using a 12T sprocket there is a loss in the power stroke of
~ 9 x 6.7%
by contrast if using a 19T sprocket the losses are closer to
~13 x 2.7%
The latter quantity is only a little over half the former.
So far as I understand it, until now, no-one has attempted to allow for this (side-)effect of chordal losses in bicycle transmissions. It might be at least (if not more) significant than the pure mechanical losses.
If so, the chances of feeling a change (because of chordal effects) that only measures about 0.5% on a test rig are likely to be higher, as indeed is the variation in real speed.
cheers