feat(SNNBot): add STDP learner and integration test (#153, #155)

- Implement reward-modulated STDP: spike timing, eligibility traces,
  reward-gated weight updates (w += η * reward * trace), clamp [-2,2]
- Add integration smoke test vs SittingDuck (compile + battle completion)
- Add tests/config.nims and nimble test task

SNNBot_garage/SNNBot.nimble

@@ -10,3 +10,6 @@ bin           = @["SNNBot"]
 requires "nim >= 2.0.0"
 requires "robocode_tankroyale_botapi >= 1.0.7"
 # radar_lock is vendored in-tree (common_libs/) and wired via config.nims --path
+
+task test, "Run integration tests":
+  exec "nim c -r --path:../common_libs tests/test_basic_battle.nim"

SNNBot_garage/src/SNNBot.nim

@@ -23,6 +23,11 @@ const
   THRESH   = 1.0         # LIF spike threshold
   MAX_GUN_TURN = 20.0    # max gun turn per tick (degrees)
   AIM_TOL  = 2.0         # arrive tolerance (degrees)
+  STDP_WIN = 20          # STDP timing window (ticks)
+  ETA      = 0.01        # learning rate
+  # ponytail: global learning rate, no per-synapse adaptation; add when performance plateaus
+  ELG_DECAY = 0.95       # eligibility trace decay per tick
+  W_CLAMP  = 2.0         # weight magnitude clamp
 
 # ── SNN types ─────────────────────────────────────────────────────────────────
 
@@ -32,12 +37,19 @@ type
     who: array[N_HID, float]               # 6×1 hidden→output weights
     vHid: array[N_HID, float]             # hidden membrane potentials
     vOut: float                            # output membrane potential (readout)
+    lastSpikeIn:  array[N_IN, int]         # tick of last input spike (-1 = never)
+    lastSpikeHid: array[N_HID, int]        # tick of last hidden spike (-1 = never)
+    eligibility:  array[N_IN * N_HID, float]  # per-synapse eligibility traces
+    tick: int                              # internal tick counter for spike timing
 
 proc initSNN(snn: var SNN) =
-  # ponytail: uniform random init, fine until STDP training added in #155
   randomize()
   for w in snn.wih.mitems: w = rand(1.0) - 0.5
   for w in snn.who.mitems: w = rand(1.0) - 0.5
+  for t in snn.lastSpikeIn.mitems:  t = -STDP_WIN - 1
+  for t in snn.lastSpikeHid.mitems: t = -STDP_WIN - 1
+  for e in snn.eligibility.mitems:  e = 0.0
+  snn.tick = 0
 
 proc encodeInput(bearing: float): array[N_IN, float] =
   ## Population-code bearing into N_IN neurons.
@@ -53,7 +65,17 @@ proc encodeInput(bearing: float): array[N_IN, float] =
 
 proc forward(snn: var SNN, inputs: array[N_IN, float]): float =
   ## One SNN tick. Returns target angle in degrees (-180..+180).
-  # Hidden layer: LIF update
+  inc snn.tick
+
+  # Decay eligibility traces each tick
+  for e in snn.eligibility.mitems: e *= ELG_DECAY
+
+  # Record input spikes
+  for i in 0 ..< N_IN:
+    if inputs[i] > 0.0:
+      snn.lastSpikeIn[i] = snn.tick
+
+  # Hidden layer: LIF update + STDP trace
   var spikes: array[N_HID, float]
   for h in 0 ..< N_HID:
     var wsum = 0.0
@@ -63,8 +85,20 @@ proc forward(snn: var SNN, inputs: array[N_IN, float]): float =
     if snn.vHid[h] >= THRESH:
       spikes[h] = 1.0
       snn.vHid[h] = 0.0
+      snn.lastSpikeHid[h] = snn.tick
+      # STDP: post fires → check recent pre-spikes (potentiation)
+      for i in 0 ..< N_IN:
+        let dt = snn.tick - snn.lastSpikeIn[i]
+        if dt >= 0 and dt <= STDP_WIN:
+          snn.eligibility[i * N_HID + h] += exp(-float(dt) / float(STDP_WIN))
     else:
       spikes[h] = 0.0
+      # STDP: pre fires after post → depression for synapses where post spiked recently
+      for i in 0 ..< N_IN:
+        if inputs[i] > 0.0:
+          let dt = snn.tick - snn.lastSpikeHid[h]
+          if dt >= 0 and dt <= STDP_WIN:
+            snn.eligibility[i * N_HID + h] -= exp(-float(dt) / float(STDP_WIN))
 
   # Output layer: membrane readout (no threshold)
   var osum = 0.0
@@ -72,8 +106,7 @@ proc forward(snn: var SNN, inputs: array[N_IN, float]): float =
     osum += spikes[h] * snn.who[h]
   snn.vOut = LEAK * snn.vOut + osum
 
-  # Scale vOut to [-180, +180]. vOut is unbounded so clamp after tanh-like squash.
-  # ponytail: simple linear scale; replace with proper output neuron tuning in #155
+  # Scale vOut to [-180, +180]
   result = tanh(snn.vOut) * 180.0
 
 # ── Bot state machine ─────────────────────────────────────────────────────────
@@ -274,6 +307,11 @@ method run*(bot: SNNBot) =
     of EVALUATE:
       let err    = abs(normalizeRelativeAngle(gunDir - bot.enemyBearing))
       let reward = 1.0 / (1.0 + err)
+      # R-STDP weight update: w += η * R * e; then clamp and decay traces
+      for idx in 0 ..< N_IN * N_HID:
+        bot.snn.wih[idx] += ETA * reward * bot.snn.eligibility[idx]
+        bot.snn.wih[idx] = bot.snn.wih[idx].clamp(-W_CLAMP, W_CLAMP)
+        bot.snn.eligibility[idx] *= ELG_DECAY
       echo "tick=" & $bot.tick & " error=" & $err & "° reward=" & $reward
       bot.phase = DECIDE