Add GPT-6.2 Sol and DeepSeek Bedrock routes to the LiteLLM gateway
The latest main push adds two Amazon Bedrock chat routes to the LiteLLM gateway and updates the runbook, observability notes, and CI expectations around them. The merged change registers the new OpenAI GPT-6.2 Sol and DeepSeek V3.2 model entries, scopes the gateway identity to declared inference permissions, and makes the staging attribution checks aware of the added routes. It is mostly a code, docs, and check change rather than a deployment change, so runtime behavior still needs verification.
Extend the gateway model registry
The gateway code adds the two Bedrock model identifiers to its model list and maps each one to a Converse request configuration. The operator documentation now asks the operator unit to verify the changed chat routes, not just LiteLLM readiness, and it records that the operator unit manages the public account model agreement for the GPT-6.2 Sol variant. The gateway identity carries inference permissions scoped to the declared models, which keeps the blast radius of a single route reviewable.
Update observability and attribution checks
The AI observability notes now describe the Bedrock model table with the new model and inference profile pairings and adjust the chat completions guidance so all routes use Converse through the regional endpoint. US inference profiles process within the United States, while DeepSeek V3.2 runs directly in Oregon. The Langfuse staging check and the LiteLLM attribution check were extended to expect the new model names and mappings, so CI can surface missing attribution or staging drift earlier. Those are code changes to the checks, and I have not run them against staging here.
Keep rollback narrow and explicit
The runbook keeps rollback as a small, auditable operation: retain the account model agreement and remove only the added route together with its inference permissions. I like that posture because a gateway facing several model vendors can turn a model level problem into a route level diff. For a platform team running multiple inference providers behind one gateway identity, the same technique gives a controlled rollback path instead of a broad model disable.