Propose removing MONGO_ALLOW_INCLUDE_INVARIANT_H (nee MONGO_INCLUDE_INVARIANT_H_WHITELISTED).
Maybe rename invariant.h to assert_util_detail.h or similar to indicate the relationship better.

The #ifdef/#error/#endif handshake in invariant.h interferes with indexing that header, as it can't be included by a trivial .cpp file to yield a successful compilation. This is mildly annoying.

The benefit it is meant to provide is hazy or maybe obsolete. It's meant to force invariant callers to include assert_util.h. But the allowlist requirement is not transitive and cannot be. Most high-level files using invariant.h without assert_util.h do so by including one of the files on the allow list (usually base/string_data.h). So they skirt the requirement, perhaps inadvertently.

I can't see what the allowlist is trying to protect. I think it's trying to produce a culture where assert_util.h will always be preferred to direct inclusion of invariant.h, which is a good thing. I think maybe renaming invariant.h to assert_util_core.h or assert_util_detail.h and retaining the admonition at the top:

/**
 * This header is separated from assert_util.h so that the low-level
 * dependencies of assert_util.h (e.g. mongo/base/status_with.h,
 * mongo/base/status.h, mongo/base/string_data.h) can use the invariant macro
 * without causing a circular include chain. It should never be included
 * directly in any other files.
 */

would do this in a more direct way that would not have any technical side effects.