Pasting here rsc's comment on the Go forum, for completeness:

I think this is a bug. Either cmds should be returned in the first, or html should not be returned in the second. Please file an issue at golang.org/issue/new. Thanks.

Labelled this as needsDecision since it needs to be decided which package to modify.

https://play.golang.org/p/DCTbO6SEruf shows that the actual problem is text's .Templates dropping all templates with no Parsed definitions. It should probably list all, like html. In any event the doc distinction is wrong.

Though I think, that undefined (not parsed) templates are still member of the association, they should be returned by text/template, not including itself can still be quite useful to simply distinguish between root and associated templates, as if you associate a template with the same name as the root template, it will overwrite the root template name in the set, in case of the html/template.

Also comparing the internal definitions I would prefer to embed text/template.Template in html/template.Template as I'm still not convinced about this comment, from html/template/template.go:22,24

	// We could embed the text/template field, but it's safer not to because
	// we need to keep our version of the name space and the underlying
	// template's in sync.

and fix the name compilation in (text/template.Template).Templates(). But I fear I cannot provide a working fix until close of 1.11

And even more: https://play.golang.org/p/ye0hz7Yqfrr or this one https://play.golang.org/p/MIFqkA1DnU4
the order of definition affects the association.

The definition of t.Templates is "Templates returns a slice of defined templates associated with t." And the definition of "defined" is "has a parsed tree associated with it". So the text/templates package is working as documented, and as I understand it should behave. I believe that html/template is the one violating the contract here.

If you believe text/template should behave differently, that's a separate argument.

Notice that the documentation in the html package is different:

// Templates returns a slice of the templates associated with t, including t
// itself.

I'm inclined to leave this all alone. The code is subtle due to the design decision to not have template sets, which I think was a mistake.

I really thought about this for a while now and I agree with your observation, that it is really hard to change the template code. On the other hand, I would like to have a smarter and more organic solution for the templates, which seems to be some work, though.
As I currently plan on writing a service, that is controlled by some sets of template data, combining text and html, I might get annoyed enough, one day, to do so.
Until then I will close this issue: The package works as documented, though the two libraries behave different, where they should do the same. The usage of templates seems to be a very important pattern, for several use cases.