Add contact name filter to SMS search tool + misc improvements

- sms search tool: accept contact name, trim/validate, skip when
  contact_id is set, pass to API client
- sms_client: new contact field in SmsSearchParams, URL-encode on wire
- Tool description clarifies contact_id takes precedence when both given
- Add parse_title_body helper for LLM response parsing
- llamacpp backend improvements

src/ai/llamacpp.rs

@@ -354,6 +354,8 @@ impl LlamaCppClient {
             .and_then(|v| v.as_i64())
             .map(|n| n as i32);
 
+        log_timings(&parsed, prompt_tokens, completion_tokens);
+
         Ok((chat_msg, prompt_tokens, completion_tokens))
     }
 }
@@ -456,6 +458,7 @@ impl LlmClient for LlamaCppClient {
             let mut role = "assistant".to_string();
             let mut prompt_tokens: Option<i32> = None;
             let mut completion_tokens: Option<i32> = None;
+            let mut last_frame: Option<Value> = None;
             let mut done_seen = false;
 
             while let Some(chunk) = byte_stream.next().await {
@@ -505,6 +508,7 @@ impl LlmClient for LlamaCppClient {
                                 .get("completion_tokens")
                                 .and_then(|n| n.as_i64())
                                 .map(|n| n as i32);
+                            last_frame = Some(v.clone());
                         }
 
                         let Some(choices) = v.get("choices").and_then(|c| c.as_array())
@@ -587,6 +591,10 @@ impl LlmClient for LlamaCppClient {
                 Some(v)
             };
 
+            if let Some(ref frame) = last_frame {
+                log_timings(frame, prompt_tokens, completion_tokens);
+            }
+
             let message = ChatMessage {
                 role,
                 content: accumulated_content,
@@ -720,6 +728,44 @@ impl LlamaCppClient {
 /// Extract a diagnostic fragment from a llama-swap / llama-server response
 /// that doesn't match the expected `{choices: [...]}` shape. llama-server
 /// returns errors as `{"error": {"message": "...", "code": N, "type": "..."}}`;
+fn log_timings(parsed: &Value, prompt_tokens: Option<i32>, completion_tokens: Option<i32>) {
+    let timings = match parsed.get("timings") {
+        Some(t) => t,
+        None => return,
+    };
+    let prompt_tps = timings.get("prompt_per_second").and_then(|v| v.as_f64());
+    let gen_tps = timings.get("predicted_per_second").and_then(|v| v.as_f64());
+    let prompt_ms = timings.get("prompt_ms").and_then(|v| v.as_f64());
+    let gen_ms = timings.get("predicted_ms").and_then(|v| v.as_f64());
+
+    let mut parts: Vec<String> = Vec::new();
+    if let Some(c) = prompt_tokens {
+        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) = completion_tokens {
+        let mut s = format!("gen={} tok", c);
+        if let Some(ms) = gen_ms {
+            s.push_str(&format!(" ({:.0} ms", ms));
+            if let Some(tps) = gen_tps {
+                s.push_str(&format!(", {:.1} tok/s", tps));
+            }
+            s.push(')');
+        }
+        parts.push(s);
+    }
+    if !parts.is_empty() {
+        log::info!("llama-swap chat metrics — {}", parts.join(", "));
+    }
+}
+
 /// llama-swap itself sometimes wraps subprocess failures with its own
 /// `{"error": "..."}` flat shape. Surface either when present, otherwise fall
 /// back to a truncated raw-JSON view.