feat(ai): streaming chat endpoint with live tool events

Add LlmClient::chat_with_tools_stream and SSE endpoint
POST /insights/chat/stream that emits text deltas, tool_call /
tool_result pairs, truncated notice, and a terminal done frame as the
agentic loop runs.

- Ollama: parses NDJSON from /api/chat stream, accumulates content
  deltas, emits Done with tool_calls from the final chunk.
- OpenRouter: parses OpenAI-compatible SSE, reassembles tool_call
  argument deltas by index, asks for stream_options.include_usage.
- InsightChatService spawns the loop on a tokio task, feeds events
  through an mpsc channel, persists training_messages at the end.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/ai/llm_client.rs

@@ -1,5 +1,6 @@
 use anyhow::Result;
 use async_trait::async_trait;
+use futures::stream::BoxStream;
 use serde::{Deserialize, Serialize};
 
 /// Provider-agnostic surface for LLM backends (Ollama, OpenRouter, …).
@@ -30,6 +31,18 @@ pub trait LlmClient: Send + Sync {
         tools: Vec<Tool>,
     ) -> Result<(ChatMessage, Option<i32>, Option<i32>)>;
 
+    /// Streaming variant of `chat_with_tools`. The returned stream yields
+    /// `TextDelta` items as content is produced, then a single terminal
+    /// `Done` carrying the complete assembled message (with tool_calls, if
+    /// any) plus token usage counts. Implementations that can't stream may
+    /// fall back to calling `chat_with_tools` and emitting the full reply
+    /// as one `Done` event.
+    async fn chat_with_tools_stream(
+        &self,
+        messages: Vec<ChatMessage>,
+        tools: Vec<Tool>,
+    ) -> Result<BoxStream<'static, Result<LlmStreamEvent>>>;
+
     /// Batch embedding generation. Dimensionality is provider/model specific.
     async fn generate_embeddings(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>>;
 
@@ -47,6 +60,25 @@ pub trait LlmClient: Send + Sync {
     fn primary_model(&self) -> &str;
 }
 
+/// Events emitted by streaming `chat_with_tools_stream`. A stream is a
+/// sequence of zero or more `TextDelta` events followed by exactly one
+/// `Done`. Callers should treat `Done` as terminal — further items (if any
+/// slip through due to upstream misbehavior) are safe to ignore.
+#[derive(Debug, Clone)]
+pub enum LlmStreamEvent {
+    /// Incremental content token(s) from the model. Concatenate in order to
+    /// reconstruct the assistant's final text.
+    TextDelta(String),
+    /// Terminal event with the full assembled message (content + any
+    /// tool_calls). `message.content` equals the concatenation of every
+    /// preceding `TextDelta.0`.
+    Done {
+        message: ChatMessage,
+        prompt_eval_count: Option<i32>,
+        eval_count: Option<i32>,
+    },
+}
+
 /// Tool definition sent to the model (OpenAI-compatible function schema).
 #[derive(Serialize, Clone, Debug)]
 pub struct Tool {