/v1/messages does not accumulate deployment tpm for native anthropic deployments

Finding: In an external report on LiteLLM main (commits 4417bf08, a 1.106.0 dev tree, and 63a4f3f2 dated 2026-10-09), non-streaming /v1/messages calls routed to a native anthropic/... deployment with deployment-level tpm set reportedly do not accumulate token usage for that deployment. Per the reporter, optional_pre_call_checks enforce_model_rate_limits therefore never returns 429 on that leg, and usage-based-routing-v2 keeps its counter at 0 (https://github.com/BerriAI/litellm/issues/45702). This is an external report, not reproduced by KFA. Reported conditions: a model group with an anthropic/claude-opus-4-8 deployment (tpm: 1) and a hosted_vllm sibling, router_settings with enforce_model_rate_limits, and a local mock upstream returning 2 tokens per call. Repeated /v1/messages calls that land on the anthropic deployment all returned HTTP 200. The same group via /v1/chat/completions reportedly returned 429 with a message that the model rate limit was exceeded (TPM limit=1, current usage=2) once the anthropic leg was hit. Moving tpm onto the OpenAI-compatible sibling reportedly triggered 429 on /v1/messages too, so the failure appears specific to the native anthropic /v1/messages path (https://github.com/BerriAI/litellm/issues/45702). Reporter's suspected cause (hypothesis, not maintainer-confirmed): the native /v1/messages response is a plain dict without _hidden_params, so response_metadata.py returns early and litellm_logging.py leaves hidden_params["litellm_model_name"] as None. The post-call increments in ModelRateLimitingCheck and LowestTPMLoggingHandler_v2 then skip or mis-key the token count (https://github.com/BerriAI/litellm/issues/45702). Related: the reporter says #41610/PR #41630 describe the same early return for streaming. Applying that PR's two library hunks locally on 63a4f3f2 reportedly left this case open, because the back-filled name is the provider-stripped claude-opus-4-8 while the pre-call key uses anthropic/claude-opus-4-8, so the counter is written under a key the check never reads (https://github.com/BerriAI/litellm/issues/45702). Status: a non-maintainer commenter said they would propagate deployment hidden params for the plain-dict path and later said a fix is up as #45720 (https://github.com/BerriAI/litellm/issues/45702#issuecomment-6091679783, https://github.com/BerriAI/litellm/issues/45702#issuecomment-6092628170). The PR was not reviewed here; its merge and release status are unknown. No verified workaround is given in the source. Users depending on deployment tpm limits should treat /v1/messages enforcement on native anthropic deployments as unreliable on the affected commits, and should check the issue and PR for updates.