LT's Silent Switcher line should be of use.
I don't think I'd suggest a custom design: even with the quirks that a lot of integrated controllers have, the fact is they offer so much more functionality than a modest sized discrete circuit can. You'll either need hundreds of components to pull it off (the last discrete design I did, used about 100 components per channel), or as many hours programming the MCU/FPGA to provide equivalent functionality in a smaller package.
I don't think you can get away from noise in general: even resonant architectures still have spiky dV or dI/dt in places, stuff like diode recovery and all that, and that will be reflected in the emissions spectrum at some level. So, expect to need filtering and shielding regardless. What you're really investigating is, how much filtering is ultimately needed, and at what frequencies?
There may be some interest in a phase interleaved design for example, which acts to reduce or null the fundamental ripple (in effect, multiplying the ripple frequency by N for N phases). That, with a Silent Switcher style control, could give quite handy improvements. That could make the difference between, say, board-level shields, versus a full shielded enclosure, or board-level filtering versus feed-thrus.
The aim is to push the ripple up to a higher frequency, so that filtering the fundamental is easier, while also bringing down the worst case harmonics, so that your filter bandwidth does not need to be insane.
Higher Fsw also helps reduce the size of components, but at the expense of higher switching losses.
It doesn't need compressed into so little spectrum as to be a pure tone; like, you could make a class E amplifier at 13.56MHz and keep everything very soft and free from harmonics, and filtering would basically amount to 13.56MHz traps (which could be a bit of space savings actually, over a lowpass of equal attenuation). Controlling something like that would really be the bigger issue though, and also the control bandwidth would fall outside of the notch filter, which really means that you don't have any filtering from load to source (i.e., the control must respond almost instantaneously to load changes).
I don't know if that kind of filtering is necessarily intended in your application (i.e., in effect, load variations being reflected as source emissions), but it may also be equivalent to the specifications you're given.
If you have to meet dropout/holdup requirements (avionics often have this requirement; yours may too?), you may need a lot of energy storage anyway, and turning that into a highly effective filter may not be much additional cost (and, cost in the most general sense, not just direct BOM cost but cost to the size or weight as well).
If size/weight is priority, it's interesting to note that the attenuation-per-energy-storage ratio (and therefore, the attenuation-per-volume, more or less) is maximized with a modest number of stages (the number depending on how much attenuation is required). That is, for a given attenuation, a 3rd order (CLC) filter would need a fairly low cutoff frequency and therefore relatively large components, but a 5th or 7th order uses more but smaller components; while a still higher order filter uses simply too many parts with size of diminishing returns (because number of components keeps rising, while cutoff frequency approaches Fstop). So there's a minima between extremes.
Tim