Fix orchestration and provider reliability

crates/galaxy_agent_core/src/provider_run.rs

@@ -98,7 +98,9 @@ impl Default for ProviderRunLimits {
     fn default() -> Self {
         Self {
             max_model_turns: 100,
-            max_model_retries_per_turn: 2,
+            // Three retries means the initial call plus three delayed retries;
+            // the fourth failure is surfaced to the user.
+            max_model_retries_per_turn: 3,
         }
     }
 }

crates/galaxy_agent_core/src/provider_run_tests.rs

@@ -930,7 +930,7 @@ fn restored_state_validation_rejects_retry_counter_and_terminal_corruption() {
     );
 
     let excessive_retries = mutate_run_json(&run(), |value| {
-        value["model_retries"] = json!(3);
+        value["model_retries"] = json!(4);
     });
     assert!(restored_state_error(&excessive_retries).contains("model-retry counter"));
 

crates/galaxy_logging/src/native.rs

@@ -513,13 +513,13 @@ fn init_internal(
 
     let mut base_logger = env_logger::builder();
 
-    base_logger.filter_level(LevelFilter::Info);
+    base_logger.filter_level(LevelFilter::Warn);
 
     // Only include `WARN` or higher logs for wgpu. By default, wgpu outputs logs at the `INFO`
     // level multiple times _per_ frame. See https://github.com/gfx-rs/wgpu/issues/3206.
     // Naga is overly noisy at `DEBUG`, so increase to `INFO`.
     base_logger
-        .filter(Some("naga"), LevelFilter::Info)
+        .filter(Some("naga"), LevelFilter::Warn)
         .filter(Some("wgpu_core"), LevelFilter::Warn)
         // Since we always pair an insertion with a deletion to avoid duplicate,
         // tantivy will log a lot of warnings for deleting a non-existing doc.