No disagreements, other than perhaps the suggestion that power is the "only" method for determining accumulated training stress, or for that matter that it's necessarily the best. All I can say is that opinions (and marketing materials) vary.
Yes, power is good for setting training at a level appropriate to current (macro) fitness, but the HRM data provides micro values that power cannot. Peak prediction isn't such grand science - periodization models have been around longer than even HRM data. It's absolutely true that power is useful for training planning, but for actual on-the-bike information, HR has some advantages.
For example, in-ride recovery times give a handy indication of what's left in the tank. If it takes too long to recover from the last hill, interval, sprint, etc., you may want to conserve for the end of the ride. Yes, you know this mostly by feel anyway, but the HR data can help separate mentally feeling like crap from actually performing like crap. For on the bike tactical decisioning, power is less useful.
Heartrate is sensitive to hydration levels, and can help keep you from overstressing - even when the power prediction is saying you should be running at a higher output, but can't know how much you are sweating.
Power (in the absence of a periodization model post-applied to the data) has no way to compensate for accumulated stress. It can say "I performed well today" or "I'm likely to perform badly this week", but it has no way to say "drop a cog, big guy." HR does, if somewhat imperfectly. One of the first concepts of any training plan is that adequate rest is as important as adequate work. Power is great for the planning of training, certainly more accurate and useful than HR data in that regard. But HR data has it's place too, and folks that ignore it are shortchanging their training plans.
I suppose what I'm trying to say is that power data talks in weeks and months, HR data in days and hours. They are complementary. Especially in the absense of adequate live-action coaching, HR gains importance as a part of training.
Interestingly, the assertion of power's value as a predictor of performance is cast into some doubt by the footnote of the article:"** I distinguish between racing and training because nine times out of ten, athletes see their personal best peak power outputs in races compared to their training. The extra adrenaline and motivation associated with competition bring outs (sic) the best data. " If power data were perfectly valid as a predictor, this would not be the case - the graph would tell all, and not all races hit all training peaks. Again, this is a macro vs micro argument, but even at the macro level there's still truth. If I'm lazy, the powermeter will say I put out less watts that day, but can't say if it's a motivational or physical issue. HR data can add info, differentiating between lazy and weak.
You've mentioned that power allows for in-training testing at VO2max, functional threshold, etc, and that's true - provided that the athlete in question actually functions at that level in that session. Two concerns here: First, plotting against inaccurate data points because the athlete isn't operating at highest functional levels that day. (HR isn't included for "historical" purposes; among other things, it's there to evaluate the quality of the power data.) Second, there exists the concern that the athlete in question will perform at those levels to create a data point during a training period that shouldn't include that level of effort. Most experts agree that this sort of testing-level effort should only occur monthly or perhaps bi-weekly. There is a strong argument that testing should be restricted to testing conditions, as much for mental as data-quality reasons. There is nothing that a powermeter provides that a speedometer on a trainer cannot also provide, other than a standardized metric, and that is convenient but not essential.
There are some theories that suggest that subpeak loads can be interpolated into adequate data points, which would make continuous power measurement more useful. But those models require a standard against which to interpolate, and that standard is - guess what - HR.
Strictly opinion: Power-based training is a terrific tool for top-level athletes with adequate coaching. Fabulous stuff for wringing out those last few percentage points of performance. For the weekend warrior thru cat3-ish racer, it's somewhere between an expensive toy and a potientially detrimental motivator. There are plenty of things that the grand or so that a powermeter costs could be spent on that would have greater long-term benefit to the athlete, including professional coaching sessions.