fix(response): unify JSON response projection across non-streaming and SSE→JSON paths

[warelik]

Fixes #2347
Related: #2046, #1396

Summary

Unifies non-streaming and forced-SSE→JSON response projection on upstream v0.5.30. Provider responses are normalized to the existing OpenAI Chat Completion pivot, then projected into the client's OpenAI, Claude, Gemini-family, Ollama, or Responses envelope. Tool calls, reasoning, finish reasons, usage, and client-native response structure are preserved instead of leaking raw OpenAI choices[] or dropping Gemini function calls.

The streaming diagnostic path emits only counts and HMAC digests. It now performs no body scans unless debug logging is explicitly enabled. Passthrough terminal handling also forwards exactly one real upstream [DONE] and synthesizes one only when upstream omitted it.

Changes

• completionProjector.js: shared OpenAI→client projection and Responses→OpenAI normalization.
• nonStreamingHandler.js / sseToJsonHandler.js: use the shared projection contract; remove duplicate/dead projection helpers.
• toolSemanticsTrace.js: safe request/provider/client tool semantics tracing.
• src/sse/utils/logger.js: export explicit debug state so tracing is not activated merely because a logger has a debug() method.
• sseDoneTracker.js: pure terminal tracker; first upstream [DONE] is forwarded, duplicates are suppressed, and flush only synthesizes a missing terminal.
• stream.js, chatCore.js, streamingHandler.js: tracing and projection wiring. Source/target trace context is carried on the existing request logger, avoiding a positional-signature conflict with passthrough-specific features.

Files

10 files: four handlers, completionProjector.js, completionProjectorSelfCheck.mjs, toolSemanticsTrace.js, sseDoneTracker.js, stream.js, and src/sse/utils/logger.js.

Verification

• completionProjectorSelfCheck.mjs: 22/22
• production build: 126/126
• built-server smoke: OpenAI and Claude streaming/non-streaming passed; exactly one OpenAI [DONE] and one Claude message_stop
• standalone git apply --check / git diff --check: clean on 9845a17

[warelik]

Code-review follow-up: logToolSemantics now early-returns when debug logging is inactive, so the hot chat/stream path skips the full request/response body scan and HMAC digest work unless the trace will actually be emitted. 20/20 self-check still passes.

[zierocode]

Fresh gap probe on 2026-07-23 at fc96ef808bb42617601f9d77113ed69decb2bcd6:

• the PR's dependency-free self-check still passes 22/22;
• Chat Completion to Responses projection omits response.model;
• it preserves token totals but drops input_tokens_details.cached_tokens and output_tokens_details.reasoning_tokens.

Reproduction:

node --input-type=module -e 'import { projectCompletionToClientFormat } from "./open-sse/translator/response/completionProjector.js"; const out=projectCompletionToClientFormat({id:"chatcmpl-gap",model:"resolved-model",choices:[{message:{role:"assistant",content:"ok"},finish_reason:"stop"}],usage:{prompt_tokens:12,prompt_tokens_details:{cached_tokens:3},completion_tokens:5,completion_tokens_details:{reasoning_tokens:2},total_tokens:17}},"openai-responses"); console.log(JSON.stringify({model:out.model,usage:out.usage,terminalShape:out.status}))'

Observed:

{"usage":{"input_tokens":12,"output_tokens":5,"total_tokens":17},"terminalShape":"completed"}

Expected additions:

{"model":"resolved-model","usage":{"input_tokens":12,"input_tokens_details":{"cached_tokens":3},"output_tokens":5,"output_tokens_details":{"reasoning_tokens":2},"total_tokens":17}}

PR #2453 already owns native Responses SSE-to-JSON preservation for these fields. This gap is specifically the unified Chat-to-client projection implemented here, so no competing projection PR is needed if it can be covered in this branch.