Streaming `reasoningContent` is exposed as accumulated full text — no incremental API for SSE typewriter UX

Background

Spring AI 2.0.1, spring-ai-starter-model-openai, streaming via ChatClient.stream().chatResponse() against an OpenAI-compatible gateway backed by DeepSeek.

Since #6373 (aligned streaming reasoning_content with non-streaming responses), every intermediate streamed chunk carries the full accumulated reasoning text in the chunk metadata (metadata["reasoningContent"]):

String accumulatedReasoning = reasoningMap.merge(id + ":" + choice.index(),
        getReasoningContent(choice), String::concat);
// ...
REASONING_CONTENT, accumulatedReasoning

I understand the rationale (the final aggregated AssistantMessage must keep the complete reasoning for advisor/chat-memory replay, since providers like DeepSeek require reasoning_content on replayed assistant tool-call messages).

Problem

Streaming the reasoning to a frontend via SSE with a typewriter effect is a first-class need (ChatGPT-style UX). But with the current API there is no incremental reasoning surface:

• Flux<ChatResponse> chunks expose only the accumulated full text under metadata["reasoningContent"] — the wire delta.reasoning_content is incremental, but it never reaches consumers.
• Consumers are forced to reconstruct deltas client-side by prefix-stripping the accumulated text (accumulate → splice → trim), which is fragile, duplicated across every app, and easy to get wrong (multi-round tool loops, stream restarts).

This mirrors #5910 (tool-call argument deltas are also swallowed by provider/framework aggregation — milestone 2.1.x).

Expected

Please provide an official incremental reasoning API, e.g. any of:

1. A per-chunk reasoningDelta field (alongside the accumulated reasoningContent) on the streamed AssistantMessage — mirrors what Vercel AI SDK does with reasoning-delta parts and what OpenAI/DeepSeek already send on the wire.
2. Or a unified stream-event API covering both reasoning deltas and tool-call arg deltas (reuse the #5910 effort).
3. At minimum, document the behavior: "intermediate chunks expose accumulated reasoning; derive deltas yourself" and show a canonical example.

Until then, every Spring AI user building a reasoning-aware chat UI has to reinvent the prefix-strip workaround.

Repro context

// 2.0.1, OpenAiChatModel stream()
chatModel.stream(prompt)
    .doOnNext(res -> {
        String full = res.getResult().getOutput().getMetadata()
                         .get("reasoningContent").toString();
        // full grows: "The" -> "The user asks" -> "The user asks what day" ...
    });

Related: #5898, #6230, #6373, #6375, #5910.

Start with OpenAiChatModel stream() and ChatClient.stream().chatResponse(), then trace how reasoningContent is placed in streamed ChatResponse metadata. Compare the related #5910 and #6373 work before choosing an incremental API shape. Done means streamed consumers can access reasoning deltas while the accumulated reasoning remains available for replay and aggregation.