evaluation extra caps litellm below 1.97.0, so no release fixing GHSA-7hp6-4w63-5g45 / CVE-2026-114612 (critical) can be installed #7252

The evaluation extra declares, in 2.7.0 (latest):

litellm<1.97.0,>=1.83.7; python_version < "3.14" and extra == "evaluation"
litellm<1.97.0,>=1.93.0; python_version >= "3.14" and extra == "evaluation"

litellm GHSA-7hp6-4w63-5g45 / CVE-2026-114612 (published 2026-10-27, critical, CVSS 3.1 9.9) affects 1.91.0 and later and is fixed in 1.100.4, 1.101.3, 1.102.2, 1.103.1 and 1.104.0rc2 and later series. Every allowed version is below 1.97.0, so the extra cannot be installed with a patched litellm on any Python: a resolver picks 1.96.2, which is in the affected range. On Python 3.14 the lower bound >=1.93.0 leaves no unaffected version at all; on older Pythons only the pre-1.91.0 releases of the 1.x google-cloud-aiplatform line (litellm<1.86.0) escape it.

Repro (Python 3.14.7, uv 0.13.4):

$ uv pip install --dry-run "google-cloud-aiplatform[evaluation]==2.7.0" "litellm>=1.100.4"
error: No solution found when resolving dependencies
  cause: Because google-cloud-aiplatform[evaluation]>=2.7.0 depends on litellm{python_full_version < '3.15'}>=1.93.0,<1.97.0 and you require google-cloud-aiplatform[evaluation]==2.7.0, we can conclude that you require litellm>=1.93.0,<1.97.0.
         And because you require litellm>=1.100.4, we can conclude that your requirements are unsatisfiable.

Downstream, apache-airflow-providers-google requires google-cloud-aiplatform[evaluation], so Airflow images on Python 3.14 ship the affected litellm and fail scanners on a critical finding.

Could the upper bound move to a range containing a patched release (1.104.x or 1.106.x), as #7233 and #7062 did for the earlier advisories? If something blocks it, what is it? I am not claiming the advisory is reachable through the evaluation client; it describes the LiteLLM proxy server.