feat(PPO_Bot): bot-relative bullets + scan staleness (STATE_DIM=57)

- Bullet state (indices 44-55): enemy-relative → bot-relative frame
  (bot needs threat vectors to itself for dodging, not to enemy)
- New index 56: scan staleness = min(ticksSinceLastScan / 30, 1.0)
  (gives policy a confidence signal for enemy data freshness)
- warm_start.py updated: 44→57 dim expansion, TARGET_DIM variable
- Tests updated for new state layout

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

PPO_Bot/network.nim

@@ -4,7 +4,7 @@ import arraymancer
 import std/[math, random]
 
 const
-  STATE_DIM*  = 56
+  STATE_DIM*  = 57
   ACTION_DIM* = 6
 
 var

PPO_Bot/state_vector.nim

@@ -1,4 +1,4 @@
-## State vector builder — produces 56-float normalized tensor for PPO policy.
+## State vector builder — produces 57-float normalized tensor for PPO policy.
 ## No bot API imports; takes plain BotState + EnemyTracker structs.
 
 import std/math
@@ -26,10 +26,11 @@ proc buildStateVector*(bot: BotStateData; enemy: EnemyTracker;
                        remainingGunAngle: float64 = 0.0;
                        bullets: openArray[BulletData] = [];
                        bulletCount: int = 0): Tensor[float32] =
-  ## Build the 56-float normalized state tensor.
+  ## Build the 57-float normalized state tensor.
   ## Indices 0-43: existing features. Indices 44-55: up to 3 bullet slots (4 floats each).
+  ## Index 56: scan staleness (ticksSinceLastScan / 30, clamped to 1).
   ## All values clipped to roughly [-1, 1] via division by physical maxima.
-  result = zeros[float32](56)
+  result = zeros[float32](57)
 
   let aW = bot.arenaWidth
   let aH = bot.arenaHeight
@@ -102,14 +103,12 @@ proc buildStateVector*(bot: BotStateData; enemy: EnemyTracker;
 
   # --- Bullet tracking (indices 44-55): up to 3 enemy bullets, 4 floats each ---
   # Per bullet: relX/aW, relY/aH, speed/20, ticksToImpact/diag
-  # Positions are relative to enemy (threat vector). Slots beyond bulletCount stay 0.
+  # Positions are relative to bot (useful for dodging). Slots beyond bulletCount stay 0.
-  let ex = if enemy.hasContact: enemy.current.x else: bot.arenaWidth  / 2.0
-  let ey = if enemy.hasContact: enemy.current.y else: bot.arenaHeight / 2.0
   for i in 0 ..< min(bulletCount, 3):
     let b = bullets[i]
     let bSpeed = 20.0 - 3.0 * b.power  # Tank Royale bullet speed formula
-    let bdx = b.x - ex
+    let bdx = b.x - bot.x
-    let bdy = b.y - ey
+    let bdy = b.y - bot.y
     let bdist = sqrt(bdx * bdx + bdy * bdy)
     let ticks = if bSpeed > 0.0: bdist / bSpeed else: 0.0
     let base = 44 + i * 4
@@ -117,3 +116,6 @@ proc buildStateVector*(bot: BotStateData; enemy: EnemyTracker;
     result[base + 1] = float32(bdy / bot.arenaHeight)
     result[base + 2] = float32(bSpeed / 20.0)
     result[base + 3] = float32(ticks / diag)
+
+  # --- Scan staleness (index 56) ---
+  result[56] = float32(min(enemy.current.ticksSinceLastScan.float64 / 30.0, 1.0))

PPO_Bot/tests/test_state.nim

@@ -84,7 +84,7 @@ block testStateVectorLength:
     arenaWidth: 800.0, arenaHeight: 600.0,
   )
   let sv = buildStateVector(bot, t)
-  check sv.shape == [56], "state vector has 56 elements"
+  check sv.shape == [57], "state vector has 57 elements"
 
 block testStateVectorRange:
   var t = initEnemyTracker()
@@ -95,7 +95,7 @@ block testStateVectorRange:
     arenaWidth: 800.0, arenaHeight: 600.0,
   )
   let sv = buildStateVector(bot, t)
-  for i in 0 ..< 56:
+  for i in 0 ..< 57:
     check sv[i] >= -2.0f32 and sv[i] <= 2.0f32,
       &"sv[{i}]={sv[i]} out of [-2,2] range"
 
@@ -161,6 +161,8 @@ block testHistoryPaddedWhenEmpty:
   # indices 44-55 (bullet slots) default to 0 when no bullets provided
   for i in 44 ..< 56:
     check sv[i] == 0.0f32, &"bullet slot {i} should be 0 when no bullets"
+  # index 56 (staleness): no contact so ticksSinceLastScan=0 → 0/30 = 0
+  check sv[56] == 0.0f32, "sv[56] staleness should be 0 when no contact"
 
 block testBulletSlots:
   var t = initEnemyTracker()
@@ -171,15 +173,15 @@ block testBulletSlots:
     arenaWidth: 800.0, arenaHeight: 600.0,
   )
   # Bullet at (300,250), power 1.0 → speed = 20-3 = 17
-  # relX = 300-400 = -100, relY = 250-300 = -50
+  # relX = 300-200 = 100, relY = 250-200 = 50 (relative to bot, not enemy)
-  # dist to enemy = sqrt(100^2+50^2) ≈ 111.8, ticks ≈ 111.8/17 ≈ 6.58
+  # dist to bot = sqrt(100^2+50^2) ≈ 111.8, ticks ≈ 111.8/17 ≈ 6.58
   let b = BulletData(x: 300.0, y: 250.0, power: 1.0)
   let sv = buildStateVector(bot, t, 0.0, 0.0, [b], 1)
-  check abs(sv[44] - (-100.0/800.0).float32) < 0.001f32, "bullet relX"
+  check abs(sv[44] - (100.0/800.0).float32) < 0.001f32, "bullet relX"
-  check abs(sv[45] - (-50.0/600.0).float32) < 0.001f32,  "bullet relY"
+  check abs(sv[45] - (50.0/600.0).float32) < 0.001f32,  "bullet relY"
   check abs(sv[46] - (17.0/20.0).float32) < 0.001f32,    "bullet speed norm"
   # second slot should be zero-padded
-  for i in 48 ..< 56:
+  for i in 48 ..< 57:
     check sv[i] == 0.0f32, &"unused bullet slot {i} should be 0"
 
 echo "All tests passed"

tools/training_runner/curriculum_retrain.sh

@@ -0,0 +1,126 @@
+#!/usr/bin/env bash
+# curriculum_retrain.sh — sequential curriculum retraining for PPO_Bot v2.
+# Bots in order: Fire → MyFirstDroid → Target → MyFirstLeader → Crazy → Corners → PaintingBot → MyFirstBot
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+PPO_BOT_SRC="$REPO_ROOT/PPO_Bot"
+
+# Source hyperparams (but we'll override opponent/rounds/log per phase)
+ENV_FILE="$SCRIPT_DIR/training.env"
+[ -f "$ENV_FILE" ] && set -a && . "$ENV_FILE" && set +a
+
+JAR="${TANK_ROYALE_JAR:-/home/davide/Projects/tank-royale/runner/examples/lib/robocode-tankroyale-runner.jar}"
+export PPO_BOT_DIR="$PPO_BOT_SRC"
+export SAMPLE_BOTS_DIR="${SAMPLE_BOTS_DIR:-/home/davide/Projects/tank-royale/sample-bots/java/build/archive}"
+
+LOGS_DIR="$SCRIPT_DIR/logs"
+SNAPSHOTS_DIR="$SCRIPT_DIR/snapshots"
+SUMMARY="$LOGS_DIR/curriculum_retrain_summary.txt"
+mkdir -p "$LOGS_DIR" "$SNAPSHOTS_DIR"
+
+ROUND_COUNTER_FILE="$PPO_BOT_SRC/weights/round_counter.txt"
+
+echo "=== PPO_Bot v2 curriculum retrain — $(date) ===" | tee -a "$SUMMARY"
+
+# ── compile once up front (run.sh does this; curriculum skipped it) ───────────
+echo ">>> Compiling PPO_Bot..."
+(cd "$PPO_BOT_SRC" && nimble build -d:release)
+echo ">>> Compiling RunTraining.java..."
+(cd "$SCRIPT_DIR" && javac -cp "$JAR" RunTraining.java)
+echo ">>> Compilation done."
+
+# ── helpers ──────────────────────────────────────────────────────────────────
+
+run_java() {
+  # run_java <opponent> <rounds>
+  local opp="$1" rounds="$2"
+  local remaining="$rounds"
+  local attempts=0
+  while [ "$remaining" -gt 0 ]; do
+    attempts=$(( attempts + 1 ))
+    [ "$attempts" -gt 1 ] && sleep 2
+    java -cp "$SCRIPT_DIR:$JAR" RunTraining "$opp" "$remaining" || \
+      echo ">>> Java exited non-zero (crash/restart expected)"
+    local done=0
+    [ -f "$ROUND_COUNTER_FILE" ] && done=$(cat "$ROUND_COUNTER_FILE" 2>/dev/null || echo 0)
+    remaining=$(( rounds - done ))
+    # Guard: if counter didn't advance at all on this attempt, something is
+    # fundamentally broken (bad binary, missing bots dir, etc.) — abort rather
+    # than loop forever.
+    if [ "$attempts" -ge 3 ] && [ "$done" -eq 0 ]; then
+      echo "ERROR: round_counter still 0 after $attempts attempts — aborting" >&2
+      exit 1
+    fi
+  done
+}
+
+count_wins() {
+  # count_wins <jsonl_file> → prints win count
+  # grep -c exits 1 on zero matches (not an error) which was triggering || echo 0
+  # making WINS="0\n0". Use awk which always exits 0.
+  local f="$1"
+  [ -f "$f" ] || { echo 0; return; }
+  awk '/"win":true/{c++} END{print c+0}' "$f"
+}
+
+snap_weights() {
+  local label="$1"
+  local counter
+  counter=$(cat "$ROUND_COUNTER_FILE" 2>/dev/null || echo 0)
+  local dst="$SNAPSHOTS_DIR/v2_${label}_r${counter}"
+  cp -a "$PPO_BOT_SRC/weights/latest" "$dst"
+  echo "  snapshot → $dst"
+}
+
+# ── curriculum ────────────────────────────────────────────────────────────────
+
+declare -A TRAIN_ROUNDS=( [Fire]=200 [MyFirstDroid]=200 [Target]=200 [MyFirstLeader]=200 [Crazy]=200
+                          [Corners]=400 [PaintingBot]=400 [MyFirstBot]=1000 )
+BOTS=(Fire MyFirstDroid Target MyFirstLeader Crazy Corners PaintingBot MyFirstBot)
+GATE=150
+GATE_PASS=148
+
+for BOT in "${BOTS[@]}"; do
+  TR="${TRAIN_ROUNDS[$BOT]}"
+  echo "" | tee -a "$SUMMARY"
+  echo "══ $BOT  (train=${TR} gate=${GATE}) ══" | tee -a "$SUMMARY"
+
+  # Reset round counter so run_java terminates correctly for each phase
+  echo 0 > "$ROUND_COUNTER_FILE"
+
+  # ── training phase ──────────────────────────────────────────────────────────
+  TRAIN_LOG="$LOGS_DIR/${BOT}_train.jsonl"
+  export TRAINING_OPPONENT="$BOT"
+  export TRAINING_ROUNDS="$TR"
+  export PPOB_LOG_FILE="$TRAIN_LOG"
+  unset PPOB_EVAL_ONLY
+  > "$TRAIN_LOG"
+  echo "  [train] $TR rounds..." | tee -a "$SUMMARY"
+  run_java "$BOT" "$TR"
+  echo "  [train] done" | tee -a "$SUMMARY"
+
+  # ── gate eval ───────────────────────────────────────────────────────────────
+  echo 0 > "$ROUND_COUNTER_FILE"
+  EVAL_LOG="$LOGS_DIR/${BOT}_eval.jsonl"
+  export TRAINING_ROUNDS="$GATE"
+  export PPOB_LOG_FILE="$EVAL_LOG"
+  export PPOB_EVAL_ONLY=1
+  > "$EVAL_LOG"
+  echo "  [eval]  $GATE rounds..." | tee -a "$SUMMARY"
+  run_java "$BOT" "$GATE"
+
+  WINS=$(count_wins "$EVAL_LOG")
+  if [ "$WINS" -ge "$GATE_PASS" ]; then
+    STATUS="PASS ($WINS/$GATE)"
+    snap_weights "$BOT"
+  else
+    STATUS="FAIL ($WINS/$GATE)"
+  fi
+  echo "  [gate]  $STATUS" | tee -a "$SUMMARY"
+  unset PPOB_EVAL_ONLY
+done
+
+echo "" | tee -a "$SUMMARY"
+echo "=== curriculum complete — $(date) ===" | tee -a "$SUMMARY"

tools/training_runner/generalist_train.sh

@@ -0,0 +1,351 @@
+#!/usr/bin/env bash
+# generalist_train.sh — adaptive mixed-opponent training for a generalist PPO_Bot.
+#
+# Problem it solves: the sequential curriculum killed earlier skills (catastrophic
+# forgetting). Instead we sample ONE opponent per chunk, weighted by how weak the
+# bot currently is against it (max(FLOOR, 1 - winrate)), and keep the best
+# generalist checkpoint by mean eval winrate across all 8 opponents.
+#
+# State / logs (all plain files, resumable):
+#   /tmp/generalist_wr.txt        — one line per opponent: "Name winrate"
+#   /tmp/generalist_seed.txt      — optional starting winrates (same format);
+#                                    wins over wr.txt when both exist
+#   /tmp/generalist_best.txt      — "mean rounds" of the best generalist eval so far
+#   /tmp/generalist_train.log     — full run log (tee'd)
+#   /tmp/generalist_eval_<Opp>.jsonl — per-opponent eval battles (fresh per eval)
+#   /tmp/generalist_train_<Opp>.jsonl — per-opponent training battles (accumulated)
+#
+# Env-overridable config: TOTAL_ROUNDS CHUNK_SIZE EVAL_INTERVAL EVAL_ROUNDS
+#   PASS_RATE FLOOR ALPHA (all declared in the config block below).
+#
+# Usage:
+#   ./generalist_train.sh            — run the full training loop
+#   ./generalist_train.sh --selftest — verify the weighted sampler against the
+#                        expected distribution (no battles, no compilation)
+#
+# Dead-bot resilience is inherited from the harness: RunTraining.java aborts
+# (exit 1) when PPO_Bot's round counter is frozen or short at battle end, and
+# this script retries; a chunk that makes no counter progress across 3 attempts
+# aborts loudly instead of silently burning rounds.
+set -uo pipefail
+export LC_ALL=C
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+PPO_BOT_SRC="$REPO_ROOT/PPO_Bot"
+
+ENV_FILE="$SCRIPT_DIR/training.env"
+[ -f "$ENV_FILE" ] && set -a && . "$ENV_FILE" && set +a
+
+JAR="${TANK_ROYALE_JAR:-/home/davide/Projects/tank-royale/runner/examples/lib/robocode-tankroyale-runner.jar}"
+export PPO_BOT_DIR="$PPO_BOT_SRC"
+export SAMPLE_BOTS_DIR="${SAMPLE_BOTS_DIR:-/home/davide/Projects/tank-royale/sample-bots/java/build/archive}"
+
+SNAPSHOTS_DIR="$SCRIPT_DIR/snapshots"
+mkdir -p "$SNAPSHOTS_DIR"
+
+ROUND_COUNTER_FILE="$PPO_BOT_SRC/weights/round_counter.txt"
+LOG="/tmp/generalist_train.log"
+WR_STATE_FILE="/tmp/generalist_wr.txt"
+WR_SEED_FILE="/tmp/generalist_seed.txt"
+BEST_FILE="/tmp/generalist_best.txt"
+SELFTEST_LOG="/tmp/generalist_sampling_selftest.log"
+
+# ── config (env-overridable) ──────────────────────────────────────────────────
+TOTAL_ROUNDS="${TOTAL_ROUNDS:-40000}"
+CHUNK_SIZE="${CHUNK_SIZE:-60}"
+EVAL_INTERVAL="${EVAL_INTERVAL:-1000}"
+EVAL_ROUNDS="${EVAL_ROUNDS:-30}"
+PASS_RATE="${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
+
+OPPONENTS=(Fire MyFirstDroid Target MyFirstLeader Crazy Corners PaintingBot MyFirstBot)
+
+# ── helpers ───────────────────────────────────────────────────────────────────
+
+# read_counter — PPO_Bot's persisted round counter; robust to a mid-write read
+# (partial/empty file → 0 instead of a garbage/arithmetic error).
+read_counter() {
+  local v=0
+  [ -f "$ROUND_COUNTER_FILE" ] || { echo 0; return; }
+  v=$(cat "$ROUND_COUNTER_FILE" 2>/dev/null || echo 0)
+  case "$v" in
+    ''|*[!0-9]*) echo 0 ;;
+    *) echo "$v" ;;
+  esac
+}
+
+count_wins() {
+  local f="$1"
+  [ -f "$f" ] || { echo 0; return; }
+  awk '/"win":true/{c++} END{print c+0}' "$f"
+}
+
+# sample_opponent — weighted draw: P(opp) ∝ max(FLOOR, 1 - winrate).
+# Cumulative distribution built in awk; random draw via awk's rand() seeded from
+# bash $RANDOM. Deterministic when $RANDOM is pre-seeded (the self-test relies on
+# this). Pass an explicit seed as $1 to override.
+sample_opponent() {
+  local seed="${1:-$RANDOM}"
+  awk -v seed="$seed" -v floor="$FLOOR" \
+      -v names="${OPPONENTS[*]}" -v rates="${WR[*]}" '
+    BEGIN {
+      srand(seed);
+      n = split(names, N, " ");
+      split(rates, R, " ");
+      total = 0;
+      for (i = 1; i <= n; i++) {
+        w = 1.0 - R[i] + 0.0;
+        if (w < floor) w = floor;
+        W[i] = w; total += w;
+      }
+      r = rand() * total;
+      acc = 0;
+      for (i = 1; i <= n; i++) {
+        acc += W[i];
+        if (r < acc) { print N[i]; exit }
+      }
+      print N[n];
+    }'
+}
+
+# run_chunk <opponent> <rounds> [train|eval] — one battle chunk, with the
+# 3-attempt retry / loud-abort guard from curriculum_retrain.sh. Prints the
+# JSONL path on success; exits the script on a zero-progress chunk.
+run_chunk() {
+  local opp="$1" want="$2" mode="${3:-train}"
+  local logfile
+  if [ "$mode" = "eval" ]; then
+    logfile="/tmp/generalist_eval_${opp}.jsonl"
+    > "$logfile"                    # fresh snapshot per eval
+    export PPOB_EVAL_ONLY=1         # freeze training (pure evaluation)
+  else
+    logfile="/tmp/generalist_train_${opp}.jsonl"
+    unset PPOB_EVAL_ONLY
+  fi
+  export TRAINING_OPPONENT="$opp"
+  export TRAINING_ROUNDS="$want"
+  export PPOB_LOG_FILE="$logfile"
+
+  # Reset the counter so RunTraining's expected-end check (start+rounds) is exact.
+  echo 0 > "$ROUND_COUNTER_FILE"
+
+  local remaining="$want" attempts=0 done=0
+  while [ "$remaining" -gt 0 ]; do
+    attempts=$(( attempts + 1 ))
+    [ "$attempts" -gt 1 ] && sleep 2
+    # RunTraining.java exits 1 on its own when the counter is frozen ("process
+    # dead") or short at battle end ("N rounds never trained ... aborting for
+    # restart") — treat that as a retry signal, then re-check the counter.
+    java -cp "$SCRIPT_DIR:$JAR" RunTraining "$opp" "$remaining" \
+      || echo ">>> Java exited non-zero (attempt $attempts/3) — crash/restart expected"
+    done=$(read_counter)
+    remaining=$(( want - done ))
+    # Zero counter progress across 3 attempts = dead bot / broken harness.
+    # The bot crashed constantly before; a crash must fail loudly now.
+    if [ "$attempts" -ge 3 ] && [ "$done" -eq 0 ]; then
+      echo "ERROR: round_counter still 0 after $attempts attempts vs $opp — aborting for restart" >&2
+      exit 1
+    fi
+  done
+  unset PPOB_EVAL_ONLY
+  echo "$logfile"
+}
+
+persist_wr() {
+  local i opp
+  : > "$WR_STATE_FILE"
+  for i in "${!OPPONENTS[@]}"; do
+    printf "%s %.4f\n" "${OPPONENTS[$i]}" "${WR[$i]}" >> "$WR_STATE_FILE"
+  done
+}
+
+load_wr() {
+  local src=""
+  if [ -f "$WR_SEED_FILE" ]; then
+    src="$WR_SEED_FILE"
+    echo ">>> Seeding winrates from $WR_SEED_FILE"
+  elif [ -f "$WR_STATE_FILE" ]; then
+    src="$WR_STATE_FILE"
+    echo ">>> Loading winrates from $WR_STATE_FILE"
+  else
+    echo ">>> No winrate state — starting all at 1.0 (uniform floor sampling until first eval)"
+  fi
+  local -A from_file=()
+  if [ -n "$src" ]; then
+    local name val
+    while read -r name val _; do
+      [ -n "$name" ] || continue
+      from_file[$name]="${val:-1.0}"
+    done < "$src"
+  fi
+  local i opp
+  for i in "${!OPPONENTS[@]}"; do
+    opp="${OPPONENTS[$i]}"
+    WR[$i]="${from_file[$opp]:-1.0}"
+  done
+}
+
+load_best() {
+  BEST_MEAN=0.0
+  BEST_ROUNDS=0
+  if [ -f "$BEST_FILE" ]; then
+    read -r BEST_MEAN BEST_ROUNDS < "$BEST_FILE" || :
+    [ -n "$BEST_MEAN" ] || BEST_MEAN=0.0
+    [ -n "$BEST_ROUNDS" ] || BEST_ROUNDS=0
+    echo ">>> Best-so-far: mean=$BEST_MEAN at round $BEST_ROUNDS"
+  fi
+}
+
+snap_weights() {
+  local rounds="$1"
+  local dst="$SNAPSHOTS_DIR/generalist_best_r${rounds}"
+  cp -a "$PPO_BOT_SRC/weights/latest" "$dst"   # fresh dir per save (rounds strictly increase)
+  echo "  saved checkpoint -> $dst"
+}
+
+# do_eval <trained_rounds> — fresh battle per opponent, PPOB_EVAL_ONLY=1,
+# EVAL_ROUNDS each. Updates winrates (eval-only!) and the best generalist
+# checkpoint by mean winrate.
+do_eval() {
+  local trained_rounds="$1"
+  local i opp w logfile oldwr evwr newwr alert
+  local total_wins=0
+  local total_possible=$(( ${#OPPONENTS[@]} * EVAL_ROUNDS ))
+  local mean
+  echo ""
+  echo "=== EVAL @ round $trained_rounds ==="
+  for i in "${!OPPONENTS[@]}"; do
+    opp="${OPPONENTS[$i]}"
+    logfile=$(run_chunk "$opp" "$EVAL_ROUNDS" eval) || { echo "  $opp: ERR"; continue; }
+    w=$(count_wins "$logfile")
+    total_wins=$(( total_wins + w ))
+    evwr=$(awk -v w="$w" -v n="$EVAL_ROUNDS" 'BEGIN{printf "%.4f", w/n}')
+    oldwr="${WR[$i]}"
+    newwr=$(awk -v a="$ALPHA" -v o="$oldwr" -v e="$evwr" 'BEGIN{printf "%.4f", a*o + (1-a)*e}')
+    WR[$i]="$newwr"
+    persist_wr   # state survives a crash mid-eval
+    alert=""
+    if awk -v e="$evwr" -v p="$PASS_RATE" 'BEGIN{exit !(e < p)}'; then
+      alert="  <-- BELOW PASS_RATE $PASS_RATE"
+    fi
+    echo "  $opp: $w/$EVAL_ROUNDS (wr $oldwr -> $newwr)$alert"
+  done
+  mean=$(awk -v n="${#OPPONENTS[@]}" -v rates="${WR[*]}" '
+    BEGIN { split(rates, R, " "); s = 0; for (i = 1; i <= n; i++) s += R[i]; printf "%.4f", s / n }')
+  echo "  mean winrate: $mean (total $total_wins/$total_possible)"
+  if awk -v m="$mean" -v b="$BEST_MEAN" 'BEGIN{exit !(m > b)}'; then
+    BEST_MEAN="$mean"
+    BEST_ROUNDS="$trained_rounds"
+    snap_weights "$trained_rounds"
+    printf '%s %s\n' "$BEST_MEAN" "$BEST_ROUNDS" > "$BEST_FILE"
+    echo "  >>> NEW BEST generalist: mean=$mean at $trained_rounds rounds"
+  else
+    echo "  best remains: $BEST_MEAN at $BEST_ROUNDS rounds"
+  fi
+}
+
+# ── self-test ─────────────────────────────────────────────────────────────────
+# Mock winrates: MyFirstBot 0.0, PaintingBot 0.1, 6 strong bots 1.0.
+# Weights: 1.0, 0.9, and 6x FLOOR 0.05 => total 2.2.
+# Expected: MyFirstBot 1.0/2.2=45.5%, PaintingBot 40.9%, each strong 2.27%.
+selftest() {
+  local rc=0 i opp d frac
+  local -A draws=()
+  WR=(1.0 1.0 1.0 1.0 1.0 1.0 0.1 0.0)   # order matches OPPONENTS
+  RANDOM=12345                            # fixed seed -> reproducible draws
+  for ((i = 0; i < 10000; i++)); do
+    opp=$(sample_opponent)
+    draws[$opp]=$(( ${draws[$opp]:-0} + 1 ))
+  done
+  echo "=== sampling self-test: 10000 draws, seed RANDOM=12345 ==="
+  echo "mock winrates: MyFirstBot 0.0, PaintingBot 0.1, strong x6 1.0"
+  echo "weights -> MyFirstBot 1.0, PaintingBot 0.9, strong 0.05 each (total 2.2)"
+  echo ""
+  echo "actual draw distribution:"
+  for opp in "${OPPONENTS[@]}"; do
+    d=${draws[$opp]:-0}
+    frac=$(awk -v d="$d" 'BEGIN{printf "%.4f", d/10000}')
+    printf "  %-14s %5d draws (share %.4f)\n" "$opp" "$d" "$frac"
+  done
+  echo ""
+  local mfb
+  mfb=$(awk -v d="${draws[MyFirstBot]:-0}" 'BEGIN{printf "%.4f", d/10000}')
+  assert_range() {
+    # assert_range <name> <frac> <lo> <hi>
+    if awk -v f="$2" -v lo="$3" -v hi="$4" 'BEGIN{exit !(f >= lo && f <= hi)}'; then
+      echo "  OK   $1 share $2 in [$3,$4]"
+    else
+      echo "  FAIL $1 share $2 not in [$3,$4]"; rc=1
+    fi
+  }
+  assert_range "MyFirstBot (weak focus)" "$mfb" 0.35 0.50
+  for opp in Fire MyFirstDroid Target MyFirstLeader Crazy Corners; do
+    frac=$(awk -v d="${draws[$opp]:-0}" 'BEGIN{printf "%.4f", d/10000}')
+    assert_range "$opp maintenance" "$frac" 0.015 0.05
+  done
+  echo ""
+  [ "$rc" -eq 0 ] && echo "SELF-TEST PASS" || echo "SELF-TEST FAIL"
+  exit "$rc"
+}
+
+# ── main training loop ────────────────────────────────────────────────────────
+main() {
+  echo "=== PPO_Bot generalist mixed training — $(date) ==="
+  echo "Config: TOTAL_ROUNDS=$TOTAL_ROUNDS CHUNK_SIZE=$CHUNK_SIZE EVAL_INTERVAL=$EVAL_INTERVAL EVAL_ROUNDS=$EVAL_ROUNDS PASS_RATE=$PASS_RATE FLOOR=$FLOOR ALPHA=$ALPHA"
+  echo "Opponents: ${OPPONENTS[*]}"
+
+  echo ">>> Compiling PPO_Bot..."
+  (cd "$PPO_BOT_SRC" && nimble build -d:release) || { echo "ERROR: PPO_Bot build failed" >&2; exit 1; }
+  echo ">>> Compiling RunTraining.java..."
+  (cd "$SCRIPT_DIR" && javac -cp "$JAR" RunTraining.java) || { echo "ERROR: RunTraining compile failed" >&2; exit 1; }
+  echo ">>> Compilation done."
+
+  load_best
+  load_wr
+  persist_wr   # materialize state now so an early crash still leaves a readable file
+
+  # fresh per-run train JSONLs; cross-chunk state lives in the wr/best files
+  local opp
+  for opp in "${OPPONENTS[@]}"; do
+    : > "/tmp/generalist_train_${opp}.jsonl"
+  done
+
+  local trained_rounds=0 chunk_rounds=0
+  while [ "$trained_rounds" -lt "$TOTAL_ROUNDS" ]; do
+    chunk_rounds=$CHUNK_SIZE
+    if [ $(( trained_rounds + CHUNK_SIZE )) -gt "$TOTAL_ROUNDS" ]; then
+      chunk_rounds=$(( TOTAL_ROUNDS - trained_rounds ))
+    fi
+    opp=$(sample_opponent)
+    echo ""
+    echo "Chunk $trained_rounds: $opp ($chunk_rounds rounds expected)"
+    run_chunk "$opp" "$chunk_rounds" train || :
+    trained_rounds=$(( trained_rounds + chunk_rounds ))
+
+    if [ $(( trained_rounds % EVAL_INTERVAL )) -eq 0 ]; then
+      do_eval "$trained_rounds"
+    fi
+  done
+
+  # final eval at the very end (skip when the loop already evaluated at TOTAL_ROUNDS)
+  if [ $(( TOTAL_ROUNDS % EVAL_INTERVAL )) -ne 0 ]; then
+    echo ""
+    echo "=== FINAL eval — $(date) ==="
+    do_eval "$trained_rounds"
+  fi
+
+  echo ""
+  echo "=== generalist training complete — $(date) ==="
+  echo "Final winrates in $WR_STATE_FILE"
+  echo "Best generalist checkpoint: $SNAPSHOTS_DIR/generalist_best_r${BEST_ROUNDS} (mean=$BEST_MEAN at $BEST_ROUNDS rounds)"
+}
+
+if [ "${1:-}" = "--selftest" ]; then
+  selftest 2>&1 | tee "$SELFTEST_LOG"
+  exit "${PIPESTATUS[0]}"
+fi
+
+main "$@" 2>&1 | tee -a "$LOG"
+exit "${PIPESTATUS[0]}"

tools/training_runner/mixed_train.sh

@@ -0,0 +1,155 @@
+#!/usr/bin/env bash
+# mixed_train.sh — mixed-opponent training for generalist PPO_Bot policy.
+# Randomly picks an opponent each chunk to prevent catastrophic forgetting.
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+PPO_BOT_SRC="$REPO_ROOT/PPO_Bot"
+
+ENV_FILE="$SCRIPT_DIR/training.env"
+[ -f "$ENV_FILE" ] && set -a && . "$ENV_FILE" && set +a
+
+JAR="${TANK_ROYALE_JAR:-/home/davide/Projects/tank-royale/runner/examples/lib/robocode-tankroyale-runner.jar}"
+export PPO_BOT_DIR="$PPO_BOT_SRC"
+export SAMPLE_BOTS_DIR="${SAMPLE_BOTS_DIR:-/home/davide/Projects/tank-royale/sample-bots/java/build/archive}"
+
+SNAPSHOTS_DIR="$SCRIPT_DIR/snapshots"
+mkdir -p "$SNAPSHOTS_DIR"
+
+ROUND_COUNTER_FILE="$PPO_BOT_SRC/weights/round_counter.txt"
+
+# ── config ────────────────────────────────────────────────────────────────────
+OPPONENTS=(Fire MyFirstDroid Target MyFirstLeader Crazy Corners PaintingBot MyFirstBot)
+TOTAL_ROUNDS=10000
+CHUNK_SIZE=60
+EVAL_INTERVAL=500
+EVAL_ROUNDS=50
+PASS_THRESHOLD=45
+
+# ── compile once ──────────────────────────────────────────────────────────────
+echo "=== PPO_Bot mixed training — $(date) ==="
+echo ">>> Compiling PPO_Bot..."
+(cd "$PPO_BOT_SRC" && nimble build -d:release)
+echo ">>> Compiling RunTraining.java..."
+(cd "$SCRIPT_DIR" && javac -cp "$JAR" RunTraining.java)
+echo ">>> Compilation done."
+
+# ── helpers ───────────────────────────────────────────────────────────────────
+
+count_wins() {
+  local f="$1"
+  [ -f "$f" ] || { echo 0; return; }
+  awk '/"win":true/{c++} END{print c+0}' "$f"
+}
+
+# run_chunk <opponent> <rounds> [eval]
+# Sets PPOB_EVAL_ONLY if third arg is "eval". Retries up to 3 times on failure.
+run_chunk() {
+  local opp="$1" rounds="$2" mode="${3:-train}"
+  local logfile
+  if [ "$mode" = "eval" ]; then
+    logfile="/tmp/${opp}_eval.jsonl"
+    export PPOB_EVAL_ONLY=1
+  else
+    logfile="/tmp/${opp}_mixed_train.jsonl"
+    unset PPOB_EVAL_ONLY
+  fi
+  export TRAINING_OPPONENT="$opp"
+  export TRAINING_ROUNDS="$rounds"
+  export PPOB_LOG_FILE="$logfile"
+  > "$logfile"
+
+  # Reset counter so java knows when it's done
+  echo 0 > "$ROUND_COUNTER_FILE"
+
+  local attempt=0
+  while [ "$attempt" -lt 3 ]; do
+    attempt=$(( attempt + 1 ))
+    [ "$attempt" -gt 1 ] && sleep 2
+    java -cp "$SCRIPT_DIR:$JAR" RunTraining "$opp" "$rounds" && break
+    echo ">>> Java exited non-zero (attempt $attempt/3)"
+  done
+
+  if [ "$attempt" -ge 3 ]; then
+    local done=0
+    [ -f "$ROUND_COUNTER_FILE" ] && done=$(cat "$ROUND_COUNTER_FILE" 2>/dev/null || echo 0)
+    if [ "$done" -eq 0 ]; then
+      echo "WARNING: chunk $opp made no progress after 3 attempts — skipping"
+      unset PPOB_EVAL_ONLY
+      return 1
+    fi
+  fi
+
+  unset PPOB_EVAL_ONLY
+  echo "$logfile"
+}
+
+evaluate_all() {
+  local trained_rounds="$1"
+  local total_wins=0
+  local total_possible=$(( ${#OPPONENTS[@]} * EVAL_ROUNDS ))
+  local scoreboard="[EVAL @ round $trained_rounds]"
+
+  for opp in "${OPPONENTS[@]}"; do
+    local logfile
+    logfile=$(run_chunk "$opp" "$EVAL_ROUNDS" eval) || { scoreboard="$scoreboard ${opp}:ERR"; continue; }
+    local wins
+    wins=$(count_wins "$logfile")
+    total_wins=$(( total_wins + wins ))
+    scoreboard="$scoreboard ${opp}:${wins}/${EVAL_ROUNDS}"
+  done
+
+  scoreboard="$scoreboard Total:${total_wins}/${total_possible}"
+  echo "$scoreboard"
+  echo "$total_wins"
+}
+
+snap_weights() {
+  local trained_rounds="$1"
+  local dst="$SNAPSHOTS_DIR/mixed_r${trained_rounds}"
+  cp -a "$PPO_BOT_SRC/weights/latest" "$dst"
+  echo "  snapshot → $dst"
+}
+
+random_opponent() {
+  local idx=$(( RANDOM % ${#OPPONENTS[@]} ))
+  echo "${OPPONENTS[$idx]}"
+}
+
+# ── training loop ─────────────────────────────────────────────────────────────
+trained_rounds=0
+best_total_wins=-1
+n_opponents=${#OPPONENTS[@]}
+total_possible=$(( n_opponents * EVAL_ROUNDS ))
+
+echo "Config: TOTAL_ROUNDS=$TOTAL_ROUNDS CHUNK_SIZE=$CHUNK_SIZE EVAL_INTERVAL=$EVAL_INTERVAL EVAL_ROUNDS=$EVAL_ROUNDS PASS_THRESHOLD=$PASS_THRESHOLD"
+echo "Opponents: ${OPPONENTS[*]}"
+
+while [ "$trained_rounds" -lt "$TOTAL_ROUNDS" ]; do
+  opp=$(random_opponent)
+  echo "--- [round $trained_rounds] training vs $opp ($CHUNK_SIZE rounds)"
+  run_chunk "$opp" "$CHUNK_SIZE" train || true
+  trained_rounds=$(( trained_rounds + CHUNK_SIZE ))
+
+  if [ $(( trained_rounds % EVAL_INTERVAL )) -eq 0 ]; then
+    # evaluate_all prints scoreboard then total_wins on separate lines
+    mapfile -t eval_out < <(evaluate_all "$trained_rounds")
+    echo "${eval_out[0]}"
+    total_wins="${eval_out[1]:-0}"
+    if [ "$total_wins" -gt "$best_total_wins" ]; then
+      best_total_wins="$total_wins"
+      snap_weights "$trained_rounds"
+    fi
+  fi
+done
+
+echo ""
+echo "=== Final evaluation — $(date) ==="
+mapfile -t eval_out < <(evaluate_all "$trained_rounds")
+echo "${eval_out[0]}"
+total_wins="${eval_out[1]:-0}"
+if [ "$total_wins" -gt "$best_total_wins" ]; then
+  snap_weights "$trained_rounds"
+fi
+echo "=== mixed training complete — $(date) ==="

tools/warm_start.py

@@ -0,0 +1,66 @@
+#!/usr/bin/env python3
+"""Warm-start TARGET_DIM weights from old 44-dim trained weights."""
+import numpy as np
+from pathlib import Path
+
+SRC = Path("tools/training_runner/snapshots/best_post_maint_r38003")
+DST = Path("PPO_Bot/weights/latest")
+OLD_DIM = 44
+TARGET_DIM = 57  # change this to expand to a different input dimension
+
+# Pad w1 [64, OLD_DIM] -> [64, TARGET_DIM] with zeros (weights and Adam moments)
+for prefix in ("actor", "critic"):
+    # Weight
+    old = np.load(SRC / f"{prefix}_w1.npy")
+    assert old.shape == (64, OLD_DIM), f"unexpected shape {old.shape}"
+    new = np.zeros((64, TARGET_DIM), dtype=old.dtype)
+    new[:, :OLD_DIM] = old
+    np.save(DST / f"{prefix}_w1.npy", new)
+    print(f"  {prefix}_w1: {old.shape} -> {new.shape}")
+
+    # Adam moments for w1: pad same way
+    adam_base = f"adam_{'a' if prefix == 'actor' else 'c'}w1"
+    for moment in ("_m", "_v"):
+        old_m = np.load(SRC / f"{adam_base}{moment}.npy")
+        new_m = np.zeros((64, TARGET_DIM), dtype=old_m.dtype)
+        new_m[:, :OLD_DIM] = old_m
+        np.save(DST / f"{adam_base}{moment}.npy", new_m)
+        print(f"  {adam_base}{moment}: {old_m.shape} -> {new_m.shape}")
+
+# Copy unchanged weight files as-is
+unchanged = [
+    "actor_w2", "actor_w3", "actor_b1", "actor_b2", "actor_b3",
+    "critic_w2", "critic_w3", "critic_b1", "critic_b2", "critic_b3",
+    "log_std",
+]
+for name in unchanged:
+    data = np.load(SRC / f"{name}.npy")
+    np.save(DST / f"{name}.npy", data)
+    print(f"  {name}: {data.shape} copied")
+
+# Copy unchanged Adam moments (all except w1, which were handled above)
+unchanged_adam = [
+    "adam_aw2", "adam_cw2",
+    "adam_aw3", "adam_cw3",
+    "adam_ab1", "adam_cb1",
+    "adam_ab2", "adam_cb2",
+    "adam_ab3", "adam_cb3",
+    "adam_logstd",
+]
+for base in unchanged_adam:
+    for moment in ("_m", "_v"):
+        data = np.load(SRC / f"{base}{moment}.npy")
+        np.save(DST / f"{base}{moment}.npy", data)
+        print(f"  {base}{moment}: {data.shape} copied")
+
+(DST / "adam_t.txt").write_text("1\n")
+(DST.parent / "round_counter.txt").write_text("0\n")
+print("  adam_t.txt -> 1, round_counter.txt -> 0")
+
+# Verify
+w1 = np.load(DST / "actor_w1.npy")
+old_w1 = np.load(SRC / "actor_w1.npy")
+assert w1.shape == (64, TARGET_DIM), f"bad shape {w1.shape}"
+assert np.allclose(w1[:, :OLD_DIM], old_w1), "old columns don't match"
+assert np.all(w1[:, OLD_DIM:] == 0), "new columns not zero"
+print(f"\nOK: actor_w1 shape={w1.shape}, cols 0-{OLD_DIM-1} match old, cols {OLD_DIM}-{TARGET_DIM-1} are zero")