Min RK

> This seems magical, and is not at all implied by allow_none=True. I don't agree with this. It's not magical to have default values for arguments. I don't see an...

It's not mutable in any meaningful way, so I don't think we should do anything about this.

It isn't usual to set `__doc__` on instances of types, is it?

What's generating that page? Is there any precedent for setting `__doc__` on instances? I'm not necessarily opposed, I'd just like to not be the first to change the meaning of...

`launch_instance` should probably return `app`. The downstream returns are such that they would inherit any changes from upstream, e.g. if it stopped being None, so that it wouldn't require changes...

I would return `app` from `launch_instance()`.

Sorry or leaving this forever. The failing tests were mainly verifying that the LazyConfig was stored so that future access would retrieve the same thing, but the asserts were broken...

I use env variables to populate traitlets in a few projects ([e.g.](https://github.com/jupyter/oauthenticator/blob/4db3cbb7d7fb5878976946c69944dba6ba0f968e/oauthenticator/oauth2.py#L86)). I've thought about a way to support it officially in traitlets, but I haven't though of a way...

I think that should work.

As described in the comment you pasted, for backward-compatibility the object must not be modified at all. The modifications in `_load_config` are modifications.