The current state of libsodium bindings for .NET is a mess. For example, here's the date of the last commit in each repo:
Besides ASodium, which was only recently developed, all of these bindings are no longer being maintained. I would also argue that it's unlikely that ASodium will be maintained in the long-term like NSec has been. Furthermore, these implementations have awful naming in my opinion and don't employ the memory protection that NSec does. Therefore, at first glance, it seems that NSec is the best libsodium binding for .NET.
However, NSec doesn't support a lot of the algorithms available in libsodium, which is a serious problem considering that it's really the only maintained binding. I completely understand why for some of them (e.g. not supporting unauthenticated ChaCha20 to encourage AEAD use, although that's unfortunate for building custom constructions), but others should be present. For instance, as briefly discussed in #27, there's no password-based KDF (aka there's no Argon2). Several of your arguments for not adding Argon2 simply don't hold up. For the parameter selection, you could just encourage the libsodium defaults and enforce sensible lower limits for the memory size and iterations. The algorithm has received wide adoption at this point, so it frankly doesn't matter whether the RFC is in the draft stage or not. Lastly, I don't see how dealing with out-of-memory exceptions is an issue when none of the other bindings seem to consider it to be a problem.
Another algorithm that should be added is XChaCha20-Poly1305, which is again widely used at this point and a more sensible option than ChaCha20-Poly1305 in many cases. I suspect your argument for not adding this is that the RFC is again in the draft stage. However, that RFC isn't very good anyway since it specifies a smaller counter than libsodium and Monocypher. Moreover, if an algorithm is accessible in libsodium, then it's probably being used, which is what's important rather than whether or not something has an RFC, especially considering how difficult it is to write an RFC and get it past the draft stage.
Then there's no secretstream support, which would save people the hassle of implementing their own stream encryption, not to mention the mistakes that often brings. There's also no crypto_box support despite that being quite popular from what I've seen, although I can understand not including this since it does rely on XSalsa20-Poly1305. There's only HKDF support rather than support for the default KDF in libsodium (salted BLAKE2b). The recommendation for generating random bytes/numbers is to use System.Security.Cryptography.RandomNumberGenerator rather than providing simple functions for libsodium, which would save people a few lines of code. Finally, there doesn't seem to be any constant time functions, and there's no padding function.
In sum, a lot of very useful algorithms in libsodium appear to be missing without justification for their absence. I can understand wanting to keep the library small, but the lack of a password-based KDF, especially when the only one available in .NET is PBKDF2, is ridiculous and so is the fact that XChaCha20-Poly1305 is missing. The presence of most of the other functions I've discussed more or less comes down to whether or not this library is trying to prevent people from making mistakes, which is implied by the 'NSec wants you to fall into the pit of success' claim.
PS, I apologise if I've made any inaccurate claims about the library. I haven't actually used NSec besides benchmarking the HKDF implementation because it doesn't include algorithms like Argon2 and XChaCha20.