fix(save_memory): bound concurrent enrichment LLM calls with semaphore

Background enrichment tasks from save_memory fired unbounded tokio::spawn
per observation, risking LLM API flooding under burst saves. Add a shared
Semaphore(3) to ObservationService that each enrichment task acquires
before proceeding.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Author: Stranmor

crates/service/src/observation_service/mod.rs

@@ -12,7 +12,7 @@ use opencode_mem_embeddings::EmbeddingService;
 use opencode_mem_llm::LlmClient;
 use opencode_mem_storage::StorageBackend;
 use opencode_mem_storage::traits::ObservationStore;
-use tokio::sync::broadcast;
+use tokio::sync::{Semaphore, broadcast};
 
 use crate::InfiniteMemoryService;
 
@@ -32,6 +32,7 @@ pub struct ObservationService {
     pub(crate) injection_dedup_threshold: f32,
     pub(crate) project_filter: Option<opencode_mem_core::ProjectFilter>,
     pub(crate) low_value_filter: opencode_mem_core::LowValueFilter,
+    pub(crate) enrichment_semaphore: Arc<Semaphore>,
 }
 
 impl ObservationService {
@@ -86,6 +87,7 @@ impl ObservationService {
             injection_dedup_threshold,
             project_filter,
             low_value_filter,
+            enrichment_semaphore: Arc::new(Semaphore::new(3)),
         }
     }
 

crates/service/src/observation_service/save_memory.rs

@@ -104,8 +104,14 @@ impl ObservationService {
         let llm = self.llm.clone();
         let storage = self.storage.clone();
         let svc = self.clone();
+        let semaphore = self.enrichment_semaphore.clone();
 
         tokio::spawn(async move {
+            let _permit = match semaphore.acquire().await {
+                Ok(permit) => permit,
+                Err(_) => return, // Semaphore closed — shut down gracefully
+            };
+
             let narrative = obs.narrative.as_deref().unwrap_or("");
             if narrative.is_empty() && obs.title.is_empty() {
                 return;