fix: request usage for NVIDIA streams

[cruush23]

Summary

• request the final usage chunk for streaming calls to NVIDIA's official OpenAI-compatible endpoint
• inject stream_options.include_usage=true only when the translated upstream payload has stream=true
• leave NVIDIA non-stream requests and other OpenAI-compatible providers unchanged

Problem

NVIDIA streaming responses return usage: null unless stream_options.include_usage is requested, so CliRelay records zero input/output tokens even though the stream succeeds.

A payload rule is not a safe workaround because payload rules match model/protocol but not the stream flag. The same rule is then applied to non-stream requests, which NVIDIA rejects with:

400 Validation: The 'stream_options' field is only allowed when 'stream' is set to true.

Implementation

The OpenAI compatibility executor now recognizes the official integrate.api.nvidia.com endpoint and adds stream_options.include_usage=true in ExecuteStream only when the final translated payload is actually streaming.

Verification

• added coverage for NVIDIA streaming injection
• added coverage that NVIDIA non-stream requests do not receive stream_options
• added coverage that other OpenAI-compatible providers remain unchanged
• go test ./internal/runtime/executor -run 'TestOpenAICompatExecutor(RequestsUsageForNVIDIAStreamsOnly|DoesNotRequestUsageForNVIDIANonStream)$' -count=1
• go test ./...
• go vet ./...
• golangci-lint run --config .golangci.yml
• go build ./cmd/server
• backend structure and restricted translator path checks passed

[cruush23]

Hi @kittors, the fork workflows are currently waiting for maintainer approval (action_required, no jobs started). I ran the repository checks locally with Go 1.26.1: focused executor tests, go test ./..., go vet ./..., golangci-lint, backend structure checks, and the server build all passed.