rtx-csm: Phase 13.5 — emotion-aware LLM prompting

Composes Phase 13.4 reactive emotion with LLM message construction so
the assistant's RESPONSE TEXT adapts to detected user tone, not just
TTS prosody.

--emotion-aware-llm flag (requires --reactive-emotion). When a non-
Neutral label is detected, the LLM-facing copy of the user message is
augmented with `\n\n[user audio tone: {label} — adjust your response
in tone and content to match]`. Per-turn only — the augmentation lives
in history_clone, never in the persistent history, so subsequent turns
aren't biased by stale signals.

Verified end-to-end: server with Q8 + LoRA + reactive-emotion +
emotion-aware-llm booted, bench turn completed 0 errors. Mock LLM
echoed back the augmented text (taking ~44s of TTS), confirming the
annotation reached the LLM. Lib suite 110/110.

The full reactive voice loop now adapts both prosody (TTS emotion_hint)
AND content (LLM annotated user message) to detected user tone. Both
paths flow through the same EmotionDetector trait, so when emotion2vec_
plus_base lands the placeholder swaps cleanly.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

crates/models/rtx-csm/examples/converse_server.rs

@@ -106,6 +106,15 @@ struct Cli {
     #[arg(long, default_value_t = false)]
     reactive_emotion: bool,
 
+    /// When `--reactive-emotion` produces a non-Neutral label, ALSO append
+    /// a per-turn signal to the LLM-facing user message so the LLM's
+    /// response text adapts (not just TTS prosody). Per turn only — the
+    /// annotation is NOT pushed to persistent history, so subsequent
+    /// turns aren't biased by stale signals. Has no effect without
+    /// `--reactive-emotion`.
+    #[arg(long, default_value_t = false)]
+    emotion_aware_llm: bool,
+
     /// Bearer token required on the WebSocket Authorization header. If
     /// unset the server is open (suitable for local dev only). Reads
     /// RTX_AUTH_TOKEN env var if not provided.
@@ -468,6 +477,11 @@ struct Shared {
     /// reactive-emotion conditioning. `emotion_hint` above is the fallback
     /// when detection yields Neutral.
     reactive_emotion: bool,
+    /// When true (and `reactive_emotion` is also true), append a per-turn
+    /// emotion-signal annotation to the LLM-facing user message so the
+    /// response text adapts in addition to the TTS prosody. Per-turn,
+    /// not pushed to persistent history (Phase 13.5).
+    emotion_aware_llm: bool,
     metrics: Metrics,
 }
 
@@ -828,6 +842,7 @@ async fn main() -> Result<()> {
         },
         emotion_hint: cli.emotion_hint.clone(),
         reactive_emotion: cli.reactive_emotion,
+        emotion_aware_llm: cli.emotion_aware_llm,
         metrics: Metrics::default(),
     });
 
@@ -1299,7 +1314,9 @@ async fn handle_connection(mut socket: WebSocket, shared: Arc<Shared>) {
         } else {
             None
         };
-        let resolved_hint = reactive_tag.or_else(|| shared.emotion_hint.clone());
+        let resolved_hint = reactive_tag
+            .clone()
+            .or_else(|| shared.emotion_hint.clone());
 
         // -- LLM + TTS: stream response back as audio chunks --------------
         history.push(ChatMessage::user(&user_text));
@@ -1327,7 +1344,30 @@ async fn handle_connection(mut socket: WebSocket, shared: Arc<Shared>) {
         // forwards to the WebSocket while concurrently watching for
         // user-audio frames (barge-in).
         let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<Vec<u8>>();
-        let history_clone = history.clone();
+
+        // Build the LLM-facing history. When --emotion-aware-llm is on AND
+        // we have a non-Neutral reactive tag (Phase 13.4), augment THIS
+        // turn's user message with a tone signal so the LLM adapts its
+        // text. The augmented message is per-turn-only — it never lands
+        // in persistent `history`, so subsequent turns aren't biased by
+        // stale signals. Phase 13.5.
+        let history_clone = if shared.emotion_aware_llm && reactive_tag.is_some() {
+            let tag = reactive_tag.as_deref().unwrap();
+            // Strip surrounding [] from the tag for the prose annotation.
+            let label = tag.trim_matches(|c| c == '[' || c == ']');
+            let mut h = history.clone();
+            if let Some(last) = h.last_mut() {
+                if matches!(last.role, rtx_csm::llm_client::Role::User) {
+                    last.content = format!(
+                        "{}\n\n[user audio tone: {label} — adjust your response in tone and content to match]",
+                        last.content
+                    );
+                }
+            }
+            h
+        } else {
+            history.clone()
+        };
         let metrics_for_tts = shared.clone();
         // Capture the moment conv_fut starts so we can derive
         // llm_to_first_audio (LLM stream + first-sentence TTS) and