tune(training): 20-round chunks, 50 eval rounds, 30k total rounds

- generalist_train.sh: CHUNK_SIZE 60→20 for faster opponent cycling
- generalist_train.sh: EVAL_ROUNDS 30→50 for more reliable eval
- training.env: TRAINING_ROUNDS→30000, removed hardcoded opponent
- warm_start.py: TARGET_DIM=57 (already committed, ensure latest)

tools/training_runner/generalist_train.sh

@@ -53,9 +53,9 @@ SELFTEST_LOG="/tmp/generalist_sampling_selftest.log"
 
 # ── config (env-overridable) ──────────────────────────────────────────────────
 TOTAL_ROUNDS="${TOTAL_ROUNDS:-40000}"
-CHUNK_SIZE="${CHUNK_SIZE:-60}"
+CHUNK_SIZE="${CHUNK_SIZE:-20}"
 EVAL_INTERVAL="${EVAL_INTERVAL:-1000}"
-EVAL_ROUNDS="${EVAL_ROUNDS:-30}"
+EVAL_ROUNDS="${EVAL_ROUNDS:-50}"
 PASS_RATE="${PASS_RATE:-0.90}"
 FLOOR="${FLOOR:-0.05}"        # minimum sampling weight — keeps maintenance sampling alive
 ALPHA="${ALPHA:-0.3}"         # winrate EMA smoothing: new = ALPHA*old + (1-ALPHA)*eval
@@ -302,6 +302,18 @@ main() {
   (cd "$SCRIPT_DIR" && javac -cp "$JAR" RunTraining.java) || { echo "ERROR: RunTraining compile failed" >&2; exit 1; }
   echo ">>> Compilation done."
 
+  # ── tmux-friendly session header ────────────────────────────────────────────
+  echo ""
+  echo "┌─────────────────────────────────────────────────────────────────┐"
+  printf "│  Start time   : %-48s│\n" "$(date '+%Y-%m-%d %H:%M:%S %Z')"
+  printf "│  Total rounds : %-48s│\n" "$TOTAL_ROUNDS"
+  printf "│  Chunk size   : %-48s│\n" "$CHUNK_SIZE  ($(( TOTAL_ROUNDS / CHUNK_SIZE )) chunks total)"
+  printf "│  Eval interval: %-48s│\n" "every $EVAL_INTERVAL rounds  ($(( TOTAL_ROUNDS / EVAL_INTERVAL )) evals planned)"
+  printf "│  Eval rounds  : %-48s│\n" "$EVAL_ROUNDS per opponent  ($(( ${#OPPONENTS[@]} * EVAL_ROUNDS )) max wins)"
+  printf "│  Opponents    : %-48s│\n" "${#OPPONENTS[@]} bots"
+  echo "└─────────────────────────────────────────────────────────────────┘"
+  echo ""
+
   load_best
   load_wr
   persist_wr   # materialize state now so an early crash still leaves a readable file

tools/training_runner/training.env

@@ -1,5 +1,5 @@
-TRAINING_OPPONENT=Fire
+# TRAINING_OPPONENT=Fire
-TRAINING_ROUNDS=3000
+TRAINING_ROUNDS=30000
 PPOB_LOG_FILE=/home/davide/Projects/SirRoboGarage/tools/training_runner/logs/fire_training.jsonl
 PPOB_LR=0.0003
 PPOB_CLIP_EPSILON=0.2