authorized_hits metric not emitted for SSE streaming responses

[noyitz]

Description

The authorized_hits metric (token count) is only emitted for non-streaming responses. When the upstream provider returns an SSE (Server-Sent Events) streaming response (e.g., Content-Type: text/event-stream), the Wasm plugin does not parse the SSE stream to extract usage.total_tokens from the final event, so the ratelimit-report to Limitador never includes the token count.

Impact

• authorized_hits{user, model, subscription} — only populated for non-streaming requests
• authorized_calls{user, subscription} — works correctly for all requests (streaming and non-streaming)
• Billing/cost dashboards based on authorized_hits undercount token usage for streaming-heavy workloads (e.g., Claude Code, OpenAI Codex which always use streaming)

Reproduction

1. Configure an ExternalModel pointing to Anthropic API with stream: true
2. Send a request through the gateway
3. Check Limitador metrics — authorized_calls increments but authorized_hits does not
4. Send the same request with stream: false — both metrics increment

Expected Behavior

The Wasm plugin should parse SSE responses to extract token usage from the final event:

• Anthropic: event: message_delta → data: {"usage": {"output_tokens": N}}
• OpenAI Responses API: event: response.completed → data: {"response": {"usage": {"total_tokens": N}}}
• OpenAI Chat Completions: final chunk with usage field when stream_options.include_usage: true

Workaround

We implemented SSE parsing in the IPP (ext_proc) response handler to extract usage data from streaming responses. See:

• Framework fix: feat: parse SSE streaming responses for response plugins llm-d/llm-d-inference-payload-processor#138
• The external-metering IPP plugin reports token counts via CloudEvents to a PostgreSQL-backed metering service

Environment

• Kuadrant Wasm shim: v0.12.1
• Limitador: v2.3.0
• OpenShift: 4.19
• MaaS: v0.1.0 (kustomize mode)