refactor(chat): remove message linearization

Remove linearize.rs module and its usage from chat preparation pipeline.

The linearization logic (merging consecutive user-like messages,
folding tool-appendable content into tool/diff messages) is no longer
needed as the underlying message processing has evolved. This
simplifies the chat preparation flow while preserving cache hits
through other mechanisms.

Includes minor fixes:
- gui: fix coin display logic in UsageCounter
- queue: handle null boost_reasoning in patches
- caps: add Default impl for CodeAssistantCaps
- session: improve skill deactivation compaction
- cache_guard: strict tools array equality for cache invalidation
- other: snapshot param handling, tool sorting, vecdb scope filtering

Author: JegernOUTT

refact-agent/engine/src/caps/caps.rs

@@ -151,6 +151,8 @@ impl ChatModelRecord {
             Some("anthropic_effort".to_string())
         } else if self.supports_thinking_budget {
             Some("anthropic_budget".to_string())
+        }  else if self.reasoning_effort_options.is_some() {
+            Some("effort".to_string())
         } else {
             None
         }
@@ -264,7 +266,7 @@ impl Default for CapsMetadata {
     }
 }
 
-#[derive(Debug, Serialize, Deserialize, Clone, Default)]
+#[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct CodeAssistantCaps {
     #[serde(deserialize_with = "normalize_string")]
     pub cloud_name: String,
@@ -303,6 +305,26 @@ pub struct CodeAssistantCaps {
     pub user_defaults: ProviderDefaults,
 }
 
+impl Default for CodeAssistantCaps {
+    fn default() -> Self {
+        Self {
+            cloud_name: String::new(),
+            telemetry_basic_dest: default_telemetry_basic_dest(),
+            telemetry_basic_retrieve_my_own: default_telemetry_retrieve_my_own(),
+            completion_models: IndexMap::new(),
+            chat_models: IndexMap::new(),
+            embedding_model: EmbeddingModelRecord::default(),
+            defaults: DefaultModels::default(),
+            caps_version: 0,
+            customization: String::new(),
+            hf_tokenizer_template: default_hf_tokenizer_template(),
+            metadata: CapsMetadata::default(),
+            model_caps: Arc::new(std::collections::HashMap::new()),
+            user_defaults: crate::providers::config::ProviderDefaults::default(),
+        }
+    }
+}
+
 fn default_telemetry_retrieve_my_own() -> String {
     "https://www.smallcloud.ai/v1/telemetry-retrieve-my-own-stats".to_string()
 }
@@ -972,9 +994,6 @@ pub async fn load_caps(
                 (caps, vec![server_provider])
             }
             Err(e) => {
-                if is_refact {
-                    return Err(format!("Cloud model catalog fetch failed: {}", e));
-                }
                 warn!("Cloud caps fetch failed ({}), falling back to local providers only", e);
                 (CodeAssistantCaps::default(), vec![])
             }

refact-agent/engine/src/chat/cache_guard.rs

@@ -67,6 +67,11 @@ fn is_append_only_prefix_inner(
         | (Value::Number(_), Value::Number(_))
         | (Value::String(_), Value::String(_)) => prev == next,
         (Value::Array(a), Value::Array(b)) => {
+            // The "tools" array is part of the prompt prefix — any change (including
+            // appending a new tool) invalidates the LLM cache. Require strict equality.
+            if parent_key == Some("tools") {
+                return a == b;
+            }
             if a.len() > b.len() {
                 return false;
             }
@@ -300,6 +305,31 @@ mod tests {
         assert!(!is_append_only_prefix(&prev, &next));
     }
 
+    #[test]
+    fn test_tools_array_strict_equality() {
+        let tool_a = json!({"type": "function", "function": {"name": "tool_a", "description": "A"}});
+        let tool_b = json!({"type": "function", "function": {"name": "tool_b", "description": "B"}});
+
+        // Identical tools → OK
+        let prev = json!({"messages": [1], "tools": [tool_a.clone()]});
+        let next = json!({"messages": [1, 2], "tools": [tool_a.clone()]});
+        assert!(is_append_only_prefix(&prev, &next));
+
+        // New tool appended mid-session → violation (breaks LLM cache prefix)
+        let next_extra = json!({"messages": [1, 2], "tools": [tool_a.clone(), tool_b.clone()]});
+        assert!(!is_append_only_prefix(&prev, &next_extra));
+
+        // Tool removed mid-session → violation
+        let prev2 = json!({"messages": [1], "tools": [tool_a.clone(), tool_b.clone()]});
+        let next_removed = json!({"messages": [1, 2], "tools": [tool_a.clone()]});
+        assert!(!is_append_only_prefix(&prev2, &next_removed));
+
+        // Tool description changed mid-session → violation
+        let tool_a_changed = json!({"type": "function", "function": {"name": "tool_a", "description": "Changed"}});
+        let next_changed = json!({"messages": [1, 2], "tools": [tool_a_changed]});
+        assert!(!is_append_only_prefix(&prev, &next_changed));
+    }
+
     #[test]
     fn test_append_only_prefix_messages_keys_strict() {
         let prev = json!({