refactor(llm): remove GBNF grammar, switch to JSON-envelope extractor

extract_content_tags now generates with grammar=None and parses the
response via a manual brace-counting JSON envelope extractor that
handles Qwen <think>...</think> prefixes and trailing stop tokens.
Five new unit tests. Bumps llama-cpp-2 to 0.1.146. Explicit
features=[] on tauri dependency (no-op).

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

crates/llm/Cargo.toml

@@ -19,7 +19,7 @@ openmp = ["llama-cpp-2/openmp"]
 magnotia-core = { path = "../core" }
 encoding_rs = "0.8"
 futures-util = "0.3"
-llama-cpp-2 = { version = "0.1.144", default-features = false }
+llama-cpp-2 = { version = "0.1.146", default-features = false }
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"

crates/llm/src/lib.rs

@@ -212,6 +212,10 @@ impl LlmEngine {
             generated.push_str(&piece);
             sampler.accept(next);
 
+            if config.grammar.is_some() && json_envelope_complete(&generated) {
+                break;
+            }
+
             if let Some(stop_index) = first_stop_index(&generated, &config.stop_sequences) {
                 generated.truncate(stop_index);
                 break;
@@ -296,13 +300,12 @@ impl LlmEngine {
     }
 
     /// 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.
+    /// pair as JSON. 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;
+    /// the parsed intent is re-validated with `is_valid_intent` and
+    /// invalid JSON bubbles as `InvalidJson` so the frontend toasts a
+    /// clear error.
     pub fn extract_content_tags(
         &self,
         transcript: &str,
@@ -338,12 +341,11 @@ impl LlmEngine {
                 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()),
+                grammar: None,
             },
         )?;
 
-        let tags: prompts::ContentTags = serde_json::from_str(raw.trim())
-            .map_err(|e| EngineError::InvalidJson(format!("{e}: raw={raw:?}")))?;
+        let tags: prompts::ContentTags = parse_json_payload(&raw)?;
         if !prompts::is_valid_intent(&tags.intent) {
             return Err(EngineError::InvalidJson(format!(
                 "intent out of closed set: {}",
@@ -459,6 +461,62 @@ fn first_stop_index(text: &str, stop_sequences: &[String]) -> Option<usize> {
         .min()
 }
 
+fn json_envelope_complete(text: &str) -> bool {
+    extract_json_envelope(text) == Some(text.trim())
+}
+
+fn extract_json_envelope(text: &str) -> Option<&str> {
+    let start = text
+        .char_indices()
+        .find_map(|(idx, ch)| (ch == '{' || ch == '[').then_some(idx))?;
+    let mut chars = text[start..].char_indices();
+    let (_, first) = chars.next()?;
+
+    let mut stack = vec![match first {
+        '{' => '}',
+        '[' => ']',
+        _ => unreachable!(),
+    }];
+    let mut in_string = false;
+    let mut escaped = false;
+
+    while let Some((offset, ch)) = chars.next() {
+        if in_string {
+            if escaped {
+                escaped = false;
+            } else if ch == '\\' {
+                escaped = true;
+            } else if ch == '"' {
+                in_string = false;
+            }
+            continue;
+        }
+
+        match ch {
+            '"' => in_string = true,
+            '{' => stack.push('}'),
+            '[' => stack.push(']'),
+            '}' | ']' => {
+                if stack.pop() != Some(ch) {
+                    return None;
+                }
+                if stack.is_empty() {
+                    let end = start + offset + ch.len_utf8();
+                    return Some(&text[start..end]);
+                }
+            }
+            _ => {}
+        }
+    }
+
+    None
+}
+
+fn parse_json_payload<T: for<'de> Deserialize<'de>>(raw: &str) -> Result<T, EngineError> {
+    let payload = extract_json_envelope(raw).unwrap_or(raw.trim());
+    serde_json::from_str(payload).map_err(|e| EngineError::InvalidJson(format!("{e}: raw={raw:?}")))
+}
+
 fn render_chat_prompt(
     model: &LlamaModel,
     messages: &[(&str, &str)],
@@ -548,6 +606,47 @@ mod tests {
         assert_eq!(index, Some(5));
     }
 
+    #[test]
+    fn json_envelope_complete_detects_finished_object() {
+        assert!(json_envelope_complete(
+            r#"{"topic":"meeting","intent":"planning"}"#
+        ));
+    }
+
+    #[test]
+    fn json_envelope_complete_detects_finished_array() {
+        assert!(json_envelope_complete(r#"["Call plumber","Buy milk"]"#));
+    }
+
+    #[test]
+    fn json_envelope_complete_ignores_braces_inside_strings() {
+        assert!(!json_envelope_complete(r#"{"topic":"literal } brace""#));
+    }
+
+    #[test]
+    fn json_envelope_complete_rejects_prefixes_and_trailing_text() {
+        assert!(!json_envelope_complete(r#"{"topic":"meeting""#));
+        assert!(!json_envelope_complete(r#"{"topic":"meeting"} extra"#));
+    }
+
+    #[test]
+    fn extract_json_envelope_skips_qwen_thinking_prefix() {
+        let raw =
+            "<think>\n\n</think>\n\n{\"topic\":\"grant-application\",\"intent\":\"planning\"}";
+        assert_eq!(
+            extract_json_envelope(raw),
+            Some("{\"topic\":\"grant-application\",\"intent\":\"planning\"}"),
+        );
+    }
+
+    #[test]
+    fn extract_json_envelope_handles_arrays_and_trailing_stop_text() {
+        assert_eq!(
+            extract_json_envelope("prefix [\"Call plumber\",\"Buy milk\"]<|im_end|>"),
+            Some("[\"Call plumber\",\"Buy milk\"]"),
+        );
+    }
+
     #[test]
     fn prompt_preflight_rejects_oversized_prompt_tokens() {
         let err = preflight_context_window(7_105, 1_024).unwrap_err();