feat(ai): chat rewind + ollama metrics logging

Rewind: POST /insights/chat/rewind truncates training_messages at a
given rendered index, dropping the target message plus any preceding
tool-call scaffolding. The initial user prompt is protected.

Metrics: log prompt_eval_count/duration and eval_count/duration from
every Ollama chat response, rendered as tokens + ms + tok/s.

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

src/ai/ollama.rs

@@ -691,6 +691,17 @@ Analyze the image and use specific details from both the visual content and the
             .await
             .with_context(|| "Failed to parse Ollama chat response")?;
 
+        // Log performance counters returned by Ollama. Durations are
+        // reported in nanoseconds; we render ms + tokens/sec for skim-ability
+        // in the server log. Missing fields are left off the line rather
+        // than printed as `None`.
+        log_chat_metrics(
+            chat_response.prompt_eval_count,
+            chat_response.prompt_eval_duration,
+            chat_response.eval_count,
+            chat_response.eval_duration,
+        );
+
         Ok((
             chat_response.message,
             chat_response.prompt_eval_count,
@@ -915,8 +926,14 @@ struct OllamaChatResponse {
     done_reason: String,
     #[serde(default)]
     prompt_eval_count: Option<i32>,
+    /// Nanoseconds spent evaluating the prompt (context ingestion).
+    #[serde(default)]
+    prompt_eval_duration: Option<u64>,
     #[serde(default)]
     eval_count: Option<i32>,
+    /// Nanoseconds spent generating the response tokens.
+    #[serde(default)]
+    eval_duration: Option<u64>,
 }
 
 #[derive(Deserialize)]
@@ -924,6 +941,52 @@ struct OllamaResponse {
     response: String,
 }
 
+fn log_chat_metrics(
+    prompt_eval_count: Option<i32>,
+    prompt_eval_duration_ns: Option<u64>,
+    eval_count: Option<i32>,
+    eval_duration_ns: Option<u64>,
+) {
+    // Compute tokens/sec when both count and duration are present.
+    fn tokens_per_sec(count: Option<i32>, duration_ns: Option<u64>) -> Option<f64> {
+        match (count, duration_ns) {
+            (Some(c), Some(d)) if c > 0 && d > 0 => Some((c as f64) * 1_000_000_000.0 / (d as f64)),
+            _ => None,
+        }
+    }
+    let prompt_ms = prompt_eval_duration_ns.map(|ns| ns as f64 / 1_000_000.0);
+    let eval_ms = eval_duration_ns.map(|ns| ns as f64 / 1_000_000.0);
+    let prompt_tps = tokens_per_sec(prompt_eval_count, prompt_eval_duration_ns);
+    let eval_tps = tokens_per_sec(eval_count, eval_duration_ns);
+
+    let mut parts: Vec<String> = Vec::new();
+    if let Some(c) = prompt_eval_count {
+        let mut s = format!("prompt={} tok", c);
+        if let Some(ms) = prompt_ms {
+            s.push_str(&format!(" ({:.0} ms", ms));
+            if let Some(tps) = prompt_tps {
+                s.push_str(&format!(", {:.1} tok/s", tps));
+            }
+            s.push(')');
+        }
+        parts.push(s);
+    }
+    if let Some(c) = eval_count {
+        let mut s = format!("gen={} tok", c);
+        if let Some(ms) = eval_ms {
+            s.push_str(&format!(" ({:.0} ms", ms));
+            if let Some(tps) = eval_tps {
+                s.push_str(&format!(", {:.1} tok/s", tps));
+            }
+            s.push(')');
+        }
+        parts.push(s);
+    }
+    if !parts.is_empty() {
+        log::info!("Ollama chat metrics — {}", parts.join(", "));
+    }
+}
+
 #[derive(Deserialize)]
 struct OllamaTagsResponse {
     models: Vec<OllamaModel>,