fix(ai): disable retry for streaming AI requests to prevent duplicate responses (#1700)

[HandSonic]

Problem

所有 AI provider（OpenAI/Claude/Gemini）使用 RetryTemplate.defaultInstance()，默认最多重试 3 次。对于流式响应，重试会从头重新发送整个流，而前端通过 accumulatedContent += newContent 简单拼接，导致 AI 回答重复出现。

Fix

将三个 provider 的 RetryTemplate.defaultInstance() 替换为 new RetryTemplate()（maxAttempts=1，不重试）。

• AiModelFactory.java:97 — OpenAI
• AiModelFactory.java:130 — Claude
• AiModelFactory.java:163 — Gemini

Related

Fixes #1700

[openai0229]

Thanks for working on this. This change does not actually disable retries. The resolved dependency is Spring Retry 2.0.12, and new RetryTemplate() initializes a SimpleRetryPolicy(3), so it still allows three attempts just like RetryTemplate.defaultInstance(). Please use an explicit policy such as RetryTemplate.builder().maxAttempts(1).build() or NeverRetryPolicy. Also note that AiModelFactory.create() is used by both chatSync and stream; if retries should only be disabled for streaming responses, the two paths need separate retry configurations.

[HandSonic]

Fixed. Replaced new RetryTemplate() with RetryTemplate.builder().maxAttempts(1).build() for all three AI providers (OpenAI, Claude, Gemini), ensuring no retries occur.

Regarding whether streaming and sync paths need separate retry configs: disabling retries uniformly in AiModelFactory.create() is reasonable, since even for sync requests, duplicate calls waste tokens and add latency. If differentiation is needed later, the two paths can be split into separate retry configurations.