[DRAFT] fix: deduplicate x-goog-api-client headers

[daniel-sanche]

We currently populate the x-goog-api-client header multiple times, in different places in the stack (api-core, google-aith, user-provided, etc). While some backend systems are able to handle the duplicate headers, it can cause issues for others

This PR explores a potential fix:

• the CapicCallable wrapper stores this header value separately at init time, and appends the combines if the user passes in a new copy of the header at runtime, instead of passing through multiple copies
• The google-auth AuthMetadataPlugin class has a new suppress_metrics_header argument. When enabled, it won't append its own x-goog-api-client header, and allow the gapic wrapper to be the only writer (this header should only be required for auth-spefici HTTP calls, not standard gapic rpcs)

b/477429588

[gemini-code-assist]

Code Review

This pull request introduces changes to handle and merge duplicate x-goog-api-client metadata headers in google-api-core and allows suppressing these metrics headers in google-auth's AuthMetadataPlugin. Feedback on the changes includes a recommendation to copy self._arbitrary_metadata to prevent shared state mutation, and a suggestion to simplify the _extract_metrics_header function into a single-pass loop for better readability and compatibility with general iterables.

[gemini-code-assist]

Assigning self._arbitrary_metadata directly to kwargs["metadata"] can lead to unexpected side effects. Since _GapicCallable is a reusable wrapper called multiple times, any downstream code or transport layer that modifies kwargs["metadata"] in-place will mutate self._arbitrary_metadata for all subsequent calls.

To prevent shared state mutation, please copy the list using list(self._arbitrary_metadata).

Suggested change

	kwargs["metadata"] = self._arbitrary_metadata
	kwargs["metadata"] = list(self._arbitrary_metadata)

[gemini-code-assist]

The current implementation of _extract_metrics_header uses index slicing (metadata[:i] and metadata[i+1:]) and multiple loops. This can be simplified into a single-pass loop that is more readable, performs better, and is more robust because it works on any iterable (including generators/iterators) without raising a TypeError on slicing.

    if not metadata:
        return [], []

    arbitrary_metadata = []
    metric_values = []
    for key, val in metadata:
        if key == client_info.METRICS_METADATA_KEY:
            metric_values.append(val)
        else:
            arbitrary_metadata.append((key, val))
    return arbitrary_metadata, metric_values