Selfplay fixes

config/default.ini

@@ -59,6 +59,8 @@ eval_bot_games = 0
 # Ladder rungs, weakest first: the [env] bot_policy ids the env's header
 # defines. Required when eval_bot_games > 0, e.g. eval_bots = 3,4,5,6.
 eval_bots = 0
+eval_bot_envs = 8192
+eval_bot_threads = 0
 
 # Per-rung [env] overrides for the bot ladder, as full section.key = value
 # lines (e.g. env.dr = 0). Env-owned so core needs no per-env knowledge.

src/constellation.c

@@ -654,7 +654,6 @@ PlotArgs DEFAULT_PLOT_ARGS = {
     .z_label = "Train/Learning Rate",
 };
 
-
 Table* dataset_table(Dataset *data, char *env) {
     for (int i = 0; i < data->n; i++) {
         if (strcmp(data->tables[i].name, env) == 0) {

src/pufferl.cu

@@ -2513,12 +2513,6 @@ typedef struct {
 #define SELFPLAY_MAX_HIST 8
 #define SELFPLAY_MAX_LADDER 16
 #define SELFPLAY_PATH_MAX 4096
-// Bot-ladder parallelism, held fixed so rung scores stay comparable across
-// sweep trials that vary vec.total_agents. selfplay.eval_bot_games / this is
-// the games each env plays; scripted bots keep their kNN across episodes, so
-// one game per env measures only cold bots.
-#define SELFPLAY_LADDER_ENVS 8192
-
 // One historical opponent ↔ policies[policy_idx] (env tag == policy_idx).
 typedef struct {
     int policy_idx;
@@ -2899,8 +2893,9 @@ static EvalResult eval_loop(Ini* ini, PuffeRL* p, int mode, int verbose,
         if (verbose) {
             puf_dashboard_print(ini, p, show, board ? epoch : 0);
         }
+        DictItem* m = dict_find(&el, metric_key);
         result.score = match ? dict_get(&el, "env/policy_0_score")
-            : dict_get(&el, metric_key);
+            : (m ? m->value : dict_get(&el, "env/score"));
         result.perf = dict_get(&el, "env/perf");
         if (match) {
             result.draw = dict_get(&el, "env/draw_rate");
@@ -3304,11 +3299,17 @@ TrainResult run_train(Ini* ini, TrainContext* ctx) {
             puf_ini_put(ini, ek, over->str);
         }
         // Fixed parallelism; ignore swept train total_agents. Each env plays
-        // bot_games / SELFPLAY_LADDER_ENVS games, so bots face a warmed-up kNN
-        // rather than being re-measured cold once per env.
+        // bot_games / ladder_envs games, so bots face a warmed-up kNN rather
+        // than being re-measured cold once per env.
         char nbuf[32];
-        snprintf(nbuf, sizeof(nbuf), "%d", SELFPLAY_LADDER_ENVS);
+        long envs = puf_ini_get(ini, "selfplay", "eval_bot_envs");
+        snprintf(nbuf, sizeof(nbuf), "%ld", envs);
         puf_ini_put(ini, "vec.total_agents", nbuf);
+        long threads = puf_ini_get(ini, "selfplay", "eval_bot_threads");
+        if (threads > 0) {
+            snprintf(nbuf, sizeof(nbuf), "%ld", threads);
+            puf_ini_put(ini, "vec.num_threads", nbuf);
+        }
         double ladder[SELFPLAY_MAX_LADDER];
         int rungs = puf_ini_get_list(ini, "selfplay", "eval_bots", ladder,
             SELFPLAY_MAX_LADDER);