fix(pipeline): handle duplicate finish_reason chunks from OpenRouter

[simon100500]

Fixes #2402

Problem

Some OpenRouter model providers (e.g. google/gemini-3.1-flash-lite-preview) send two consecutive SSE chunks with finish_reason: "tool_calls". The second chunk arrives after streamingToolCallParser.reset() has already been called, so it carries empty parts — no functionCall entries.

handleChunkMerging treated every finish chunk as authoritative and overwrote pendingFinishResponse with the empty duplicate, discarding the functionCall parts correctly assembled from the first finish chunk.

This caused processStreamResponse to see hasToolCall=false and throw:

Model stream ended with empty response text.

Fix

In handleChunkMerging: when a second finish chunk arrives and a pendingFinishResponse already exists, only merge usageMetadata (if present) and keep the candidates from the first finish chunk.

if (isFinishChunk) {
  if (hasPendingFinish) {
    // Duplicate finish chunk — keep candidates from first, merge only metadata
    const lastResponse = collectedGeminiResponses[...];
    if (response.usageMetadata) lastResponse.usageMetadata = response.usageMetadata;
    setPendingFinish(lastResponse);
  } else {
    collectedGeminiResponses.push(response);
    setPendingFinish(response);
  }
  return false;
}

Testing

The existing pipeline.test.ts suite should cover regressions. A new test case can be added for the duplicate-finish-chunk scenario if desired.

[Mingholy]

Thanks for the contribution!
This is a core change, and has some conflicts with #2404. I'm merging them into a single test branch to validate. This may take some time and will be merged after the validation!