Log avg_distance_per_infraction (port from emerge/temp_training)

Adds the metric:
  avg_distance_per_infraction = total_distance_travelled / total_infractions
which tracks how far agents drive between offroad/collision/red-light
events — a useful single-scalar driving-quality signal for wandb.

The two underlying log fields already exist in puffer-4 and are already
aggregated per-step in add_log; only the binding-side ratio was missing.

Both fields are normalized per-agent by static_vec_aggregate_logs, but
the ratio is invariant to that 1/n scaling so we just compute it
directly. Tiny epsilon clamp on the denominator guards against
div-by-zero in the rare case of zero infractions across the entire
aggregation window.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Eugene Vinitsky

sim/binding.c

@@ -183,6 +183,14 @@ void my_init(Env *env, Dict *kwargs) {
 }
 
 void my_log(Log *log, Dict *out) {
+    // Log fields are normalized per-agent by static_vec_aggregate_logs, but
+    // a ratio of two such fields is invariant to that 1/n scaling, so the
+    // per-agent rates give the same answer as the cross-population totals.
+    // Tiny epsilon clamp guards against div-by-zero when no agent had an
+    // infraction in the entire aggregation window.
+    float avg_distance_per_infraction =
+        log->total_distance_travelled / fmaxf(1e-6f, log->total_infractions);
+
     dict_set(out, "score", log->score);
     dict_set(out, "episode_return", log->episode_return);
     dict_set(out, "episode_length", log->episode_length);
@@ -191,6 +199,7 @@ void my_log(Log *log, Dict *out) {
     dict_set(out, "num_goals_reached", log->num_goals_reached);
     dict_set(out, "avg_speed_per_agent", log->avg_speed_per_agent);
     dict_set(out, "dnf_rate", log->dnf_rate);
+    dict_set(out, "avg_distance_per_infraction", avg_distance_per_infraction);
     dict_set(out, "n", log->n);
 }