feat(phase9): LlmEngine::extract_content_tags + smoke test

Added as a method on LlmEngine alongside cleanup_text and
extract_tasks; same render_chat_prompt -> generate -> parse pattern.
Truncates the transcript to its trailing 2000 chars on a UTF-8 char
boundary, runs at temperature 0.0 with the CONTENT_TAGS_GRAMMAR GBNF,
and re-validates intent against INTENT_CLOSED_SET to catch the
unlikely grammar bypass case. max_tokens 96 is enough for the JSON
envelope. Smoke test gated on KON_LLM_TEST_MODEL like the existing
smoke.rs.

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

crates/llm/src/lib.rs

@@ -285,6 +285,64 @@ impl LlmEngine {
         self.extract_tasks_with_feedback(transcript, &[])
     }
 
+    /// Phase 9 content-tag extraction. Emits a single (topic, intent)
+    /// pair under the `CONTENT_TAGS_GRAMMAR` GBNF. Truncates to the
+    /// trailing 2000 chars of the transcript so the prompt budget
+    /// stays well under any model's context window. Determinism is
+    /// enforced by temperature 0.0 and the closed-set intent grammar
+    /// rule; on the rare case the model emits a parse-able-but-out-of-
+    /// set intent, we re-validate with `is_valid_intent` and bubble
+    /// `InvalidJson` so the frontend toasts a clear error.
+    pub fn extract_content_tags(
+        &self,
+        transcript: &str,
+    ) -> Result<prompts::ContentTags, EngineError> {
+        if transcript.trim().is_empty() {
+            return Err(EngineError::Inference("empty transcript".into()));
+        }
+
+        // Truncate to the last 2000 chars on a UTF-8 char boundary so
+        // we don't slice through a multi-byte sequence.
+        const MAX_CHARS: usize = 2000;
+        let tail = if transcript.len() > MAX_CHARS {
+            let mut adj = transcript.len() - MAX_CHARS;
+            while adj < transcript.len() && !transcript.is_char_boundary(adj) {
+                adj += 1;
+            }
+            &transcript[adj..]
+        } else {
+            transcript
+        };
+
+        let model = self.loaded_model_arc()?;
+        let prompt = render_chat_prompt(
+            &model,
+            &[
+                ("system", prompts::CONTENT_TAGS_SYSTEM),
+                ("user", &format!("Transcript:\n{tail}")),
+            ],
+        )?;
+        let raw = self.generate(
+            &prompt,
+            &GenerationConfig {
+                max_tokens: 96,
+                temperature: 0.0,
+                stop_sequences: vec!["<|im_end|>".to_string(), "<|im_end_of_text|>".to_string()],
+                grammar: Some(grammars::CONTENT_TAGS_GRAMMAR.to_string()),
+            },
+        )?;
+
+        let tags: prompts::ContentTags = serde_json::from_str(raw.trim())
+            .map_err(|e| EngineError::InvalidJson(format!("{e}: raw={raw:?}")))?;
+        if !prompts::is_valid_intent(&tags.intent) {
+            return Err(EngineError::InvalidJson(format!(
+                "intent out of closed set: {}",
+                tags.intent,
+            )));
+        }
+        Ok(tags)
+    }
+
     /// Feedback-conditioned variant of `extract_tasks`. See
     /// `decompose_task_with_feedback` for the `examples` semantics.
     pub fn extract_tasks_with_feedback(