fix(openai): project aggregated citations to top-level field instead of leaking transformer_metadata - #2572

Problem

When chat completion streams carry provider citations (Perplexity /
OpenRouter style, top-level citations on each chunk),
AggregateStreamChunks collects them into
llm.Response.TransformerMetadata["citations"] and marshals the unified
llm.Response directly as the response body. The wire shape therefore
contains the internal envelope:

{"id":"chatcmpl-123",...,"transformer_metadata":{"citations":["https://..."]}}

and no top-level citations — while the streaming chunks themselves and
the direct non-streaming path (ResponseFromLLM, which extracts citations
from TransformerMetadata into the top-level field) both expose
top-level citations. The TransformerMetadata field is documented as
"ignored when serializing to JSON and is only used internally by
transformers"; the aggregator is the one place that marshals the unified
response verbatim, so the internal envelope leaks to clients.

Two caller-visible consequences:

1. A non-streaming client on a stream-required channel
   (autoAggregateStream) receives citations in a shape no OpenAI client
   reads — the citations are effectively lost.
2. The persisted request-log body for any streaming citation response
   records the internal envelope instead of the API shape; re-parsing it
   through ToLLMResponse (which reads top-level citations) loses them.

Fix

Keep collecting and deduplicating citations as today, then marshal a thin
client-facing wrapper that projects them back to the top-level
citations field. The internal envelope is never serialized. Streams
without citations produce byte-for-byte identical output.

Note on the existing test

TestAggregateStreamChunks_WithCitations previously asserted the leaked
shape (citations under transformer_metadata after a JSON round-trip).
It pinned the bug rather than a contract: the field's own documentation
says it is never serialized, and ResponseFromLLM is the codebase's
sanctioned unified→wire projection, which places citations top-level. The
test now asserts the corrected shape; its content assertions are
unchanged.

Testing

• New aggregator_citations_test.go: citations spread across chunks are
  deduplicated, sorted, and emitted top-level; the body contains no
  transformer_metadata; content/finish_reason/usage intact. A second
  test guards that citation-free streams keep byte-identical shape. The
  main test fails on the base commit with the leaked envelope visible in
  the output.
• go test ./transformer/openai/... — pass.
• go test ./... (llm module) — pass except the pre-existing
  Windows-only
  TestWebSocketStreamReturnsErrorWhenReusedConnectionClosesBeforeEvent
  failure in openai/responses, which fails identically on the clean
  base and is unrelated to this change.
• go vet ./... — clean.

Summary by CodeRabbit

• Bug Fixes
  • Aggregated streaming responses now include deduplicated, sorted citations in a top-level citations field when citations are present.
  • Responses without citations omit the citations field.
  • Aggregation continues to preserve the response message, finish reason, and usage information, so adding citations does not change those parts of the response.