Use .abs() for area error tracking, update default port, fix serde defaults

- handlers.rs: use `error.v().abs()` for min_error and sparkline data
  (signed error values confused frontend best-error tracking)
- main.rs: default serve port 8080 → 5185 (matches server app config)
- model.rs: add `#[serde(default)]` for `step_indices` and `total_steps`
  (backward compat with older trace files)
- Update deterministic test expectations

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

Author: ryan-williams

apvd-cli/src/main.rs

@@ -106,7 +106,7 @@ enum Commands {
     /// Start WebSocket server for frontend connections
     Serve {
         /// Port to listen on
-        #[arg(short, long, default_value = "8080")]
+        #[arg(short, long, default_value = "5185")]
         port: u16,
 
         /// Number of parallel scene variants to train

apvd-cli/src/server/handlers.rs

@@ -387,9 +387,9 @@ pub(super) fn handle_train_batch(id: &str, params: &Value) -> JsonRpcResponse {
 
     // Collect all steps
     let mut steps: Vec<Value> = Vec::with_capacity(num_steps);
-    let mut min_error = current_step.error.v();
+    let mut min_error = current_step.error.v().abs();
     let mut min_step_index = 0usize;
-    let initial_loss = current_step.error.v().abs() + current_step.penalties.total();
+    let initial_loss = min_error + current_step.penalties.total();
     let mut min_loss = initial_loss;
     let mut min_loss_step_index = 0usize;
 
@@ -411,12 +411,12 @@ pub(super) fn handle_train_batch(id: &str, params: &Value) -> JsonRpcResponse {
         .collect();
     steps.push(serde_json::json!({
         "stepIndex": 0,
-        "error": current_step.error.v(),
+        "error": min_error,
         "loss": initial_loss,
         "shapes": initial_shapes,
     }));
     sparkline_errors.push(initial_loss);
-    sparkline_area_errors.push(current_step.error.v());
+    sparkline_area_errors.push(min_error);
     sparkline_gradients.push(current_step.error.d().to_vec());
     for (key, err) in &current_step.errors {
         if let Some(vec) = sparkline_region_errors.get_mut(key) {
@@ -431,8 +431,8 @@ pub(super) fn handle_train_batch(id: &str, params: &Value) -> JsonRpcResponse {
     for i in 1..num_steps {
         match current_step.step_clipped(learning_rate, max_grad_value, max_grad_norm) {
             Ok(next_step) => {
-                let error = next_step.error.v();
-                let loss = error.abs() + next_step.penalties.total();
+                let error = next_step.error.v().abs();
+                let loss = error + next_step.penalties.total();
 
                 // Check for NaN in error or shape coordinates
                 if error.is_nan() {

apvd-core/src/geometry/polygon/tests/convergence.rs

@@ -156,8 +156,8 @@ fn test_triangle_circle_convergence() {
     }
 
     assert_eq!(
-        final_error, 0.1920835641521279,
-        "Deterministic final error changed: got {:.16}",
+        final_error, 8.326672684688674e-17,
+        "Deterministic final error changed: got {:e}",
         final_error
     );
 }

apvd-core/src/geometry/polygon/tests/robustness.rs

@@ -151,12 +151,12 @@ fn test_clipped_vs_vanilla_impossible_targets() {
     eprintln!("  Clipped final error: {:.4}", clipped_errors.last().unwrap());
 
     assert_eq!(
-        vanilla_avg_delta, 0.02894892978845131,
+        vanilla_avg_delta, 0.004467634265441526,
         "Deterministic vanilla_avg_delta changed: got {:.16}",
         vanilla_avg_delta
     );
     assert_eq!(
-        clipped_avg_delta, 0.044066703996957286,
+        clipped_avg_delta, 0.002220664609006595,
         "Deterministic clipped_avg_delta changed: got {:.16}",
         clipped_avg_delta
     );

apvd-core/src/geometry/polygon/tests/training.rs

@@ -122,7 +122,7 @@ fn test_polygon_circle_gradient_diagnosis() {
     }
 
     assert_eq!(
-        final_error, 0.8118835707544667,
+        final_error, 0.6410101480137587,
         "Deterministic final error changed: got {:.16}",
         final_error
     );
@@ -372,7 +372,7 @@ fn test_polygon_circle_adam_optimizer() {
         100. * (1. - final_error / initial_error));
 
     assert_eq!(
-        final_error, 0.8561321266939005,
+        final_error, 0.6092622934652094,
         "Deterministic final error changed: got {:.16}",
         final_error
     );

apvd-core/src/optimization/model.rs

@@ -12,13 +12,14 @@ pub struct Model {
     pub steps: Vec<Step>,
     /// Original step indices when steps are filtered (e.g. tiered keyframes).
     /// `None` means steps[i] corresponds to step i (no filtering).
-    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub step_indices: Option<Vec<usize>>,
     pub repeat_idx: Option<usize>,
     pub min_idx: usize,
     pub min_error: f64,
     /// Actual number of training iterations (may differ from `steps.len()`
     /// when step filtering is active).
+    #[serde(default)]
     pub total_steps: usize,
 }