fix(streaming): comply with OpenAI usage / stream_options spec

[localai-bot]

Summary

Closes #8546 and continuedev/continue#10113.

LocalAI emitted "usage":{"prompt_tokens":0,...} on every streamed chunk because OpenAIResponse.Usage was a value type without omitempty. Consumers of the official OpenAI Node SDK (continuedev/continue, Kilo Code, Roo Code, Zed, IntelliJ Continue) filter on a truthy result.usage to detect the trailing usage chunk — so the zero-but-non-null usage on every intermediate chunk made them swallow every content chunk and surface an empty chat response while the server log looked successful.

I confirmed the bug against a live LocalAI instance by replaying Continue's filter logic — 16/16 content chunks were swallowed.

Changes

• core/schema/openai.go — Usage *OpenAIUsage with omitempty so intermediate chunks no longer carry a usage key. Add OpenAIRequest.StreamOptions mirroring OpenAI's request field with include_usage.
• core/http/endpoints/openai/chat.go, completion.go — keep Usage on the struct as an in-process channel for the running cumulative, but strip it before JSON marshal. When the request set stream_options.include_usage: true, emit a dedicated trailing chunk with "choices": [] and the populated usage — matching the OpenAI spec and llama.cpp's server.
• chat_emit.go — new streamUsageTrailerJSON helper; drop the usage parameter from buildNoActionFinalChunks since chunks no longer carry usage.
• image.go, inpainting.go, edit.go — wrap Usage values with & for the new pointer field.
• UI — send stream_options:{include_usage:true} from the React (core/http/react-ui/src/hooks/useChat.js) and legacy (core/http/static/chat.js) chat clients so the token-count badge keeps populating now that the server is spec-compliant.

How LocalAI compares now

	Before	After
usage on every chunk	yes ({0,0,0})	no
Honors stream_options.include_usage	no	yes
Final choices:[] trailer chunk	no	yes (when opted in)
Continue / OpenAI-SDK consumers	drops all content	works
llama.cpp server parity	no	yes

Tests

• New chat_stream_usage_test.go (Ginkgo) pins the spec invariants:
  • intermediate chunks have no usage key
  • the trailer JSON has "choices":[] and a populated usage
  • OpenAIRequest parses stream_options.include_usage
• chat_emit_test.go updated to reflect that finals no longer embed usage.

Continue's filter logic, replayed against the new shape: 4/5 chunks yielded, 1 trailer carrying the usage totals — exactly what the OpenAI SDK expects.

Test plan

• CI: go test ./core/schema/ ./core/http/endpoints/openai/ ./core/http/middleware/ (passing locally)
• Manual: stream a chat completion from VS Code + Continue against a build of this branch; UI should show streamed content (was empty before)
• Manual: confirm React UI token-count badge still populates after the change

🤖 Generated with Claude Code