fix(spring-ai): emit final response on stream completion instead of punctuation heuristic

[caps-xia]

Link to Issue or Description of Change

Closes: #1608

Problem:

MessageConverter#toLlmResponse classifies each streaming chunk as partial/final via isPartialResponse, an ASCII terminal-punctuation heuristic (., !, ?, \n). Two failure modes:

• A turn ending with CJK terminal punctuation (。！？) is never classified final → the response is never persisted to the session, and the next turn has no memory of what the model said.
• Conversely, a mid-stream chunk that happens to end with . fires a premature final inside a longer answer.

There is also a related defect: since Runner streams partial events but never persists them (by design — see the comment in Runner), and the punctuation-"final" chunk only carries its own delta text, sessions today persist only the last punctuation-ending chunk fragment instead of the full answer.

Solution:

Stop classifying chunks mid-stream. Streaming chunks are always partial; StreamingResponseAggregator (already present but never wired) produces the single aggregated final response on stream completion:

• SpringAI streaming path: feed each converted chunk to a per-request aggregator and forward it; on the Flux completion callback, emit aggregator.getFinalResponse() — exactly one final event carrying the full text and accumulated tool calls — before onComplete().
• MessageConverter: streaming chunks are always partial; the old punctuation heuristic is retained (unused) for rollback/reference.
• The aggregated final carries the stream's last non-empty usageMetadata, which would otherwise be lost with the unpersisted partial events.

Why the completion signal: it is the only end-of-turn signal guaranteed across providers — finishReason spelling/availability varies (stop/end_turn/STOP; tool-call turns use tool_calls; some gateways omit it entirely), and punctuation is locale-dependent. This also matches the framework's intended design: Runner already expects "the final aggregated event" to carry the full content, and the Gemini bridge (Gemini.java) uses the same partials + aggregated-final pattern via its own StreamingResponseAggregator.

Behavioral notes:

• Persistence timing in the correct case is unchanged: the final chunk and stream completion are the same moment.
• Tool-call turns: accumulated function calls ride on the aggregated final, so execution timing is unchanged (tool calls arrive with the last chunks of a turn).
• Mid-stream errors now produce no final, so truncated turns are no longer persisted as if complete.

Testing Plan

Unit Tests:

• testGenerateContentStreamingCjkTerminalPunctuationPersistsFinal — a CJK-terminated turn produces exactly one final (persisted) event with the full text.
• testToLlmResponseStreamingAlwaysMarksChunksPartial — streaming chunks are always partial regardless of punctuation.
• testGenerateContentStreamingFinalCarriesUsageMetadata — usage metadata from the last chunk survives onto the aggregated final.
• Updated testGenerateContentStreaming and testGenerateContentStreamingBackpressure for the final event (3 chunks + 1 final; 1000 + 1).
• Full module suite passes locally: mvn -pl contrib/spring-ai -am test → 217 tests, 0 failures, BUILD SUCCESS.

Manual End-to-End (E2E) Tests:

Reproduced with a real OpenAI-compatible gateway and a Chinese-language workflow:

• Before: responses ending with 。！？ never persisted to the session (next turn lost context); non-CJK turns persisted only the last punctuation-ending fragment.
• After: every turn persists exactly once with the full text; live partial rendering is unaffected.

Checklist

• I have read the CONTRIBUTING.md document.
• My pull request contains a single commit.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules. — N/A, no dependent changes.

Additional context

Implements option 2 as discussed on the issue (root fix — stream-completion-based final detection; @hirematha confirmed the reproduction there). isPartialResponse is kept unused as rollback/reference documentation of the previous behavior. Happy to also provide the interim CJK-punctuation patch as a separate PR if maintainers want a short-term mitigation.

[hirematha]

Hi @caps-xia, thank you for your contribution and quick fix.It is currently under review by our team. We will update you if any additional details are needed.