There are two significant misconceptions from your colleague:
- Type 2 dielectrics, in general, need voltage derating, or some [fixed] ratio thereof.
- All ceramic capacitors, in general, need derating.
The underlying reasons I can only guess at, of course, but my guess would be they erroneously generalized #2 from #1: that is, applying the same principle, without understanding the reasons why #1 is done (or isn't) in the first place.
The reasoning behind #1, in turn, is not stated (and doesn't really matter here), so it may be that it is done correctly -- but it may also be that it is done erroneously as well. I've seen this more than a few times myself, so it seems worth discussing again.
Type 2 dielectrics exhibit nonlinearity, that is, capacitance varies with applied voltage. It also varies quite strongly with temperature. The type code describes the temperature characteristic. Voltage dependence is not defined by any standard (at least, that I know of, or at least not these tempco standards), and is up to the manufacturer to represent, and the user to select.
To emphasize: voltage rating does not determine electrical characteristics.
You can equally well find a 1uF 50V X7R 0805 that is -90% capacitance at rated voltage, or -70%. (You probably won't find one that's -30%; though I don't recall offhand what the highest I've seen is.) Not to mention, the height of the chip can vary, and further still, the amount of active material -- the manufacturer is free to put electrodes throughout the full stack height of the chip, or just a few in the middle, as long as it meets breakdown voltage ratings at a high enough yield to be economical to sell.
The takeaway from all this is, an electrically large chip, for its volume, is likely to be a tall (e.g. height = width) chip, that fills the whole volume with electrodes; but is also likely to have a substantial C(V) effect, i.e. C(Vmax) maybe -90% of C(0). Smaller values, in a given chip size, can have fuller C(V) curves (more valued retained under bias), but can also be made the "cheap" way, and you have no way of knowing short of measuring the thing.
What does this mean for derating?
In general, we cannot pick some fixed ratio. There are tons of capacitors that are perfectly serviceable at rated voltage. Most 6.3V chips I've seen are this way; at least at modest values like 10uF 0805. Conversely, I've seen "1uF" 250V X7R 2220s that are close to -95% at rated.
Perhaps they simply can't make layers thin enough to break down at very low voltages -- we're talking 100s nm layer thickness here, and ~thousands cm2, so a manufacturing defect (pinhole, crack, etc.) is far more likely than avalanche breakdown, and corresponding C(V) reduction while approaching it. Which is probably also why the C(V) / C(0) can be so much worse at high voltages.
What should we do, then?
The only thing we can do, is set an acceptable lower limit (for which we should also include basic value tolerance, change with temperature, and aging effect), find the corresponding point on the C(V) curve, and keep shopping until we find acceptable parts.
Tedious, yes -- I often take ten or twenty minutes to select a new capacitor by this process. It's unfortunate that manufacturers don't provide tools to search their parts automatically, but it's rare enough they provide such data at all.
As for C0G and other type 1 dielectrics -- they do not exhibit a C(V) dependence, so can be used at (or indeed perhaps beyond*) ratings, with no derating.
Other Limits
Mind that physical limitations may apply, for example an 0805 capacitor might be available up to 1kV, but it sure as hell ain't gonna count for 1kV as, say, "basic" insulation; functional at best, but even then it may have to be potted for reliable operation. (On that note, being ceramic material, they do have good CTI for insulation purposes; trouble is, one face is always in contact with a poor-CTI circuit board...) So, mind realistic voltage ratings, and choose body size/type accordingly. There are 1808 size chips for this purpose, for example.
Or, for risk items like cracking, where you might want to use series parts in case one can fail shorted -- obviously, the remaining part needs to handle the full voltage, even though the series pair nominally handles double the voltage. Cracking can be mitigated with choice of termination material, or by managing the mechanical design more closely (more rigid assembly, keep capacitors away from flex points).
Overvoltage
*At your own risk, of course.
Whether transient (or continuous, for that matter) overvoltage is acceptable in a design, is a different matter. If you need to compromise reliability to trade for cost, space, etc., overvoltage may be an option. You'll most likely want to work closely with the manufacturer(s) to determine what likely failure and aging statistics are, and so, how much voltage is acceptable in your application.
Note that ceramics don't fail incrementally, like self-healing film capacitors do. They fail shorted, one and done. So, first of all, your risk curve is a lot steeper to use ceramics this way.
Overvoltage can be used responsibly. Famously, compact fluorescent lamps (CFLs) operate with a starting capacitor, that sinks a large current through the lamp filaments to heat them up and begin conduction (thermionic emission), in the process developing a substantial voltage. A (physically) small capacitor is used to save cost: these types actually provide failure statistics in the datasheet. They are chosen so that, on average, they fail (value decreases below operational tolerance) about as often as the tube itself does. (Well, probably not exactly, but in the ballpark.) These are well engineered devices -- or at least, they can be, where "well" includes exploring these darker, statistically motivated sides of engineering!
Note that parts are tested to some overvoltage during production (in a suitable method, lot sampling or total). Real parts may break down at several times rated voltage -- how many, you just don't know, of course. I've heard of 16V X7Rs handling anywhere from 30 to 120V, for example.
In contrast, self-healing film capacitors don't fail shorted (or at least, are unlikely to), but as the name suggests, they evaporate the affected region and continue at reduced ratings (ESR goes up and C goes down). The change is tiny at first, but as healing events stack up, eventually whole sections of electrode become weakly connected or broken entirely, and ESR and C diverge rapidly. This progressive failure mode is well suited to EMI filters (exposed to occasional mains surge), and CFLs (startup cycles are ultimately limited by the lamp itself), hence their selection.
Also, pay attention to the difference between AC RMS (if given) and DC ratings. Usually the DC rating is closer to the peak-to-peak rating. Some manufacturers specify conditions for converting between these; usually DC is the same as AC + bias (no zero crossing), i.e. the peak voltage matters; and AC (zero crossings, or no DC) counts extra (use p-p voltage). Additional derating may apply, and, of course, check current or power ratings, or maximum AC/ripple.
