fix(vision): make streaming opt-in and bounded

Preserve the existing non-streaming request shape by default while allowing callers to opt into SSE, a total request budget, and single-attempt behavior. Reuse the shared retry policy and HTTP error handling, reject empty analyses, bound response buffers, and report incomplete streams as truncated.

Document the compatibility controls and cover streaming, JSON fallback, timeout, UTF-8 chunking, and empty-response paths.

Co-authored-by: qiuYliangM <185303122+qiuYliangM@users.noreply.github.com>
Signed-off-by: hexin <he.xin@h3c.com>

Author: h3c-hexin

crates/tui/src/config.rs

@@ -2967,6 +2967,19 @@ pub struct VisionModelConfig {
     /// Base URL for the vision model API. Defaults to OpenAI.
     #[serde(default)]
     pub base_url: Option<String>,
+    /// Total request budget in seconds, including retries and response body
+    /// consumption. `None` or `0` preserves the 120-second default; larger
+    /// values are capped at one hour.
+    #[serde(default)]
+    pub request_timeout_secs: Option<u64>,
+    /// Whether to request an SSE response. `None` preserves the existing
+    /// non-streaming request shape; callers must opt in with `Some(true)`.
+    #[serde(default)]
+    pub stream: Option<bool>,
+    /// Whether the existing transient-error retry policy is enabled. `None`
+    /// preserves the current default; `Some(false)` makes one attempt.
+    #[serde(default)]
+    pub retry_on_transient_errors: Option<bool>,
 }
 
 /// `[runtime_api]` table — knobs for the local HTTP/SSE daemon.