feat(simulation): voice pipeline integration — Factual bypass, observer wiring, tell iteration (#650 #652 #651)

ContentType::Factual bypasses LLM for fact-bearing lines (numbers,
denials). Observer enrichment systems rewrite dialogue/conversation
text with voiced versions before snapshot assembly. Friendly and
RoutineDeviation tells get concrete surface-pattern examples.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

server/src/voice/worker.rs

@@ -16,7 +16,7 @@ use crossbeam_channel::Receiver;
 
 use crate::npc::blueprint::CultureProfile;
 use crate::voice::cache::{CacheKey, VoiceCacheStore};
-use crate::voice::prompt_builder;
+use crate::voice::prompt_builder::{self, ContentType};
 use crate::voice::queue::VoiceRequest;
 
 /// Minimum token count for a valid response. Below this, retry once.
@@ -326,6 +326,13 @@ fn process_request(
     request: &VoiceRequest,
     ctx: &WorkerContext,
 ) {
+    // Factual lines bypass the LLM entirely — serve base text directly.
+    // 2B models corrupt numbers and invert denials (Spike 2 finding, D-138).
+    if request.content_type == ContentType::Factual {
+        cache_base_text(request, ctx);
+        return;
+    }
+
     let culture = match ctx.cultures.get(&request.culture_id) {
         Some(c) => c,
         None => {