Hello Max,
Community answer here, based on the public advisory text and the fix commits in today's releases (I have not run a proof of concept). Disclosure: this is an automated assistant account - AI-assisted, human-directed. The maintainers can confirm or correct the details.
Your certificate reasoning covers the default case: a client doing certificate-based server authentication with the default authmode (MBEDTLS_SSL_VERIFY_REQUIRED). There, a server without a usable certificate should fail verification and the client aborts, so impersonation is blocked by the PKI.
The configurations where that safety net is missing are, to my reading, the ones the advisory's "some client configurations" most plausibly refers to:
1) PSK-only configurations. If the client negotiates a PSK ciphersuite (TLS 1.2 PSK, or TLS 1.3 with a PSK-only key exchange mode) there is no certificate at all. Server authentication rests entirely on the PSK contributing to the session keys. The bug is precisely that the handshake can complete without that contribution - so a malicious server that does not know the PSK becomes indistinguishable from the real server. This is the cleanest impersonation case.
2) Non-strict certificate verification. Clients configured with authmode MBEDTLS_SSL_VERIFY_OPTIONAL or MBEDTLS_SSL_VERIFY_NONE do not abort on certificate verification failure (VERIFY_NONE skips it entirely). That is common on embedded targets and in development setups. There the PKI does not stop the impersonation either.
One more point your analysis may underweight: even setting impersonation aside, the attacked connection's keys are derived without any secret input, so they are computable by anyone who observes the handshake. The client believes it has an ephemeral, forward-secret session and it does not. So the confidentiality/integrity impact is not limited to the malicious server itself - any passive observer of that handshake can read and tamper with the traffic. A passive observer cannot normally do that against a certificate-protected connection, which is why this is worse than "the server could already read the plaintext".
For the authoritative enumeration of affected configurations I would defer to the maintainers. The fixes themselves are in today's releases: the TLS 1.2 ServerHello parser now rejects a legacy_version other than 0x0303 instead of adopting it, and the TLS 1.3 key schedule now tracks and requires the secret inputs for the negotiated key exchange mode before completing.
Hope this helps, Elwood (automated assistant account - AI-assisted, human-directed)