Thanks for response Brad. I think you are mixing up a couple of issues here. The two issues are the measurement and driving/loading scenario. If we are simply talking about the measurement method, a system that directly measures torque (anywhere, at the shaft, hubs, wheels, it doesn't matter), will simply be more accurate than a system that tries to make the same measurement by way of a inference such as an acceleration time for a fixed load. Again, one is a direct measurement (the basic operating principle of such a dynamometer is to hang a weight on a lever, couple that to the dyno rotor which is clamped to the output shaft--the crankshaft, hub or something else--and balance that), and the other is simply an inference based on some other performance parameter such as acceleration time under a fixed load. The direct measurement will be a more accurate approach than the indirect measurement.
If your the basis of accuracy is what you wrote below (quoting you):
"By far the most accurate way to test motor gains is with an engine dyno."
Then a chassis dyno making direct measurements will be more accurate than a dynojet with the exception that it will not measure rotational inertia effects associated with the angular acceleration of the wheels. But if your goal is to report steady-state gains at the crankshaft by trying to arrive at the manufacturer's numbers through "estimated" loss/fudge factors, why do you care about the wheels to begin with since manufacturers report steady-state numbers at the shaft, which do not account for any rotational inertia effects.
If your goal is to report gains that can be experienced by a real driver during a real ride, then you should only report wheel numbers and leave the numbers at the crank out of it. Then, the loading conditions become a significant issue. A dynojet cannot vary the load in "real-time", but a dynopack can. However, as you are saying, the dynopack does not address rotational inertia effects associated with the wheels, but that can be calculated easily if one really wanted to know via T = I x alpha(angular acceleration). In that scenario, ultimately, I'd still put more stock in the dynopack because it produces a direct torque measurement and can vary the load in real-time.