Ahem... I didn't assume a control scheme. I also didn't mention the complications/benefits of (usually loosely) coupling the two inductors. In fact, I didn't mention a lot of things in a jaunty little forum post... 
It's the magnitude that implies the assumption -- if it were say 1/3, that might be a typical placement for the common mode resonance; 1/100th can basically only be with filter components included, and 1/5th again of that implies a voltage mode control.
I can't actually think of any other justification for such a constraint right now?
Well, the RHPZ is usually what drives the crossover frequency that low rather than the actual control scheme, but upon reviewing my old project files (this was from 2013) it seems I used the infernal TI TPS40211 controller IC (which, AFAICT, is just the AEC-qualified version of the damnable TPS40210). Back then I thought this was a great and versatile chip, but after using it a few more times over the subsequent years I have stricken it - and any other TPS chip - forever from consideration because of intractable instability that I suspect is from an overly noise-sensitive current sense input. Hence that generic admonition I gave to drop the crossover frequency to 1/500th of the switching frequency might have more to do with the misery I experienced with this chip rather than with the SEPIC in general.
But yes, that is something you'd expect with a VMC converter and/or really trying to stay away from a RHPZ.
That said, CMC might only reliably eliminate one inductor from the transfer function (theoretically you can take both out at the same time, but this is a case where theory and practice don't align too well).
CMC..?
Whose theory? If it doesn't align well, it must not be very good.
Sorry - CMC = Current Mode Control; VMC = Voltage Mode Control.
But as for theory vs. reality, a favorite example of mine is the dawning realization in the late 90's that injecting a little bit of the clock ramp into the current sense input of a CMC controller can improve its noise tolerance at light loads and eliminate subharmonic oscillation at heavy loads. Eventually theory worked out the reasons for using so-called slope compensation, but only after the fact, not before it.
That said VMC with input feed-forward and pulse-by-pulse current limiting is a worthy contender to CMC with slope-compensation, particularly when dealing with wide input voltage and load ranges (CMC only indirectly achieves input FF as a result of the slope of the current sense signal changing with input voltage).
...the SEPIC it is quite difficult to avoid LC resonances between Cc and the two inductors, even more difficult if the inductors are coupled...
Hmm, I've had zero problem with them, treating them as the usual (e.g. peak or average current mode flyback) and not being particularly careful with the coupling.
Is it because I never use uncoupled inductors? Seems superfluous to me; I think I would go for buck-boost (inverting) or "flying inductor" (Vin <> Vout) if I can't get a coupled inductor of adequate rating.
No, same here - I only used off-the-shelf coupled inductors (IIRC, Bourns makes a series with intentionally high leakage for SEPICs). Then again, I only have a sample size of 2 SEPIC designs, and one of those was modestly modified from an LT app note so doesn't really count.
Have I been playing too fast and loose with these? I haven't sat down and done an analysis, and a 4th order transfer function sounds pretty awful. (Though I'm pretty sure that should be 4th with a near total pole-zero cancellation, so it's overall more like 2nd order, like the rest?)
Well, there's two LC networks to contend with (input L, coupling C, output L, output C) and the RHPZ when in boost mode so it seems true 4th order behavior would be difficult to avoid even with CMC. However, I haven't given it much thought, either, since the one real SEPIC I did design was part of a much larger system so I didn't really have the time to mess around with it. Drastically lowering f
c was an expedient, if not exactly elegant/correct, way of solving a stability problem that might or might not have been because of the transfer function.