To guide consistent behavior across languages when user attempts to set a user header of the same key, this could be to override an existing version that ships with client library, or attempting to add this header when server does not opt in for this feature.
Let's state something along the lines of:

I'd prefer not to add that, personally. There's only so much you can do to protect users who are determined to do weird things. The chances of someone accidentally setting that header are very slim - and if someone really wants to do so, they can always send a request directly from an HTTP client. We definitely shouldn't provide any way of specifically overriding the header - but I don't think "code that allows headers to be specified" should be modified to try to prevent this.

There's only so much you can do to protect users who are determined to do weird things.

I agree on this. On the other hand, it also seems valuable to give user usable feedback if we don't intend have this overriden.

To add a bit context in Java, we already have some logic trying to resolve conflict between internal header set by defaults and user headers. I am inclined to add logic to throw exception when user attempt to add a header with key "x-goog-api-version" (draft pr). Alternative to this is no change to this existing logic, which will throw exception only when internal header has this key (that's when service has a api version). I don't feel too strongly, but the first one gives a bit more consistent feedback.

I realize this is only dealing with an edge case, but also feel that this case is language agnostic, so wanted to check if we want to deal with it consistently across languages.

WDTY?

I think if Java wants to attempt to stop the user from doing it, that's fine - I don't think it should be in the AIP.

I'd prefer not to add that, personally. There's only so much you can do to protect users who are determined to do weird things. The chances of someone accidentally setting that header are very slim - and if someone really wants to do so, they can always send a request directly from an HTTP client. We definitely shouldn't provide any way of specifically overriding the header - but I don't think "code that allows headers to be specified" should be modified to try to prevent this.

I think we still need the empty [] for an implicit link.

Also let's format the annotation name as code, might require updating the definition text as well.

Applies here and throughout.

Don't use requirement words in non-requirement contexts

Do we need to hyperlink every use of google.api.api_version?

I'm guessing this is a requirement?

I does! A few more tweaks thanks Tony :)

So these first two sentences would be really good in the guidance section e.g.

Clients **must** treat the value of `gogle.api.api_version as opaque. Specifically, the value **must not**
be interpreted or parsed in any way.

Then this subsection could start:

Treating the `google.api.api_version` value as opaque is important to ensuring robust
compatibility guarantees. By relying on....

Fixed in a228cba