fix(responses): report truncation and failure in streaming conversion - #2171

Problem

Two gaps in the Responses streaming conversion made abnormal terminations look like success.

1. Upstream Responses -> Chat Completions client (outbound_stream.go): response.completed always produced finish_reason: stop (or tool_calls), ignoring the status field on the response. A response truncated by max_output_tokens was reported to the client as a clean stop; a failed response was reported as a normal completion.

2. Upstream Chat Completions -> Responses client (inbound_stream.go): the final response.completed was always built with status: completed. A stream that ended with finish_reason: length (truncation) or content_filter (moderation rejection) reached the Responses client as a success.

Changes

• outbound_stream.go: map response.completed status onto finish_reason:
  • incomplete -> length, or content_filter when incomplete_details.reason says so
  • failed -> error
  • cancelled/canceled -> cancelled
  • absent / completed status keeps the existing stop/tool_calls inference
• inbound_stream.go: map finish_reason onto the response status before building response.completed:
  • length -> incomplete with incomplete_details.reason: max_output_tokens
  • content_filter -> incomplete with reason: content_filter
  • error -> failed
  • cancelled/canceled -> cancelled
  • other values leave the default completed
• Both places that previously set status: completed unconditionally now only do so when no terminal status was mapped, so the finish-reason mapping survives both the usage-chunk path and the stream-end fallback.

Tests

• TestOutboundTransformer_TransformStream_MapsCompletedStatusToFinishReason: incomplete / content_filter / failed / cancelled / canceled / nil status / completed + tool calls
• TestInboundTransformer_TransformStream_MapsFinishReasonToCompletedStatus: length / content_filter / error / cancelled / canceled / unknown / stop / tool_calls
• TestInboundTransformer_TransformStream_UsageBeforeFinishReasonKeepsMappedStatus: usage chunk arriving before the finish_reason chunk

go test ./llm/transformer/openai/responses passes (the websocket executor test fails locally on Windows with a wsarecv error unrelated to these changes; it passes in CI).

Summary by CodeRabbit

• Bug Fixes

  • Improved stream status handling so completion, incomplete, failed, and cancelled responses are preserved correctly.
  • Added accurate finish reasons for length limits, content filtering, errors, cancellations, and tool calls.
  • Prevented usage updates from overwriting previously determined terminal statuses.

• Tests

  • Added coverage for inbound and outbound status and finish-reason mappings, including usage events received before completion details.