I forgot about the replaygain_peak value, that indeed is the easy way to quantify the effect. Yes, it is much more important than I expected. Still, it is limited: if my math is correct, the signal in your example should be attenuated by 1.72 dB to avoid clipping. The track replaygain is -3.65, and thus is more than enough to prevent clipping.
Scanning my collection for some loud recordings (-8, -9 dB replaygain), I see that the album peak value (which would be the loudest peak in the collection) rarely exceeds 1.3 (+2.3 dB) (wait, found a nasty one with album peak of 1.45, replay gain -10 dB). For older not remastered albums with replay gain values in the order of -2, -3 dB, I find album peak values to be at most just above one (e.g. 1.01, 1.07). This would still lead me to conclude that the issue is of practical importance for dynamic tracks only, these requiring a replay gain value of only -2 dB and up (compared to the 89 dB reference level, and as far as they are really dynamic, i.e. have also peaks near the maximum).
Because of the magnitude of potential clipping on decoding, I would be even more inclined to opt for the "Don't clip" option except where loudness rather than quality is important (such as when on the road or in the car).
QUOTE
Anyway, the second link pointed by carpman should have been enough.
Is it? It was a conceptual example, but in no way a real world, quantitative example. Thank you for pointing me to the real extent of the effect anyway.