fix(gun_harness): min 15 obs before competing, window 50→100, KNN min k=5

Gun selector now gates competition: guns with <15 total observations across
all power bins sit out until at least one gun qualifies. Falls back to ungated
selection if no gun reaches threshold, preventing cold-start stalls.

Sliding window increased from 50 to 100 ticks to reduce switching noise.

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

common_libs/gun_harness/virtual_bullets.nim

@@ -8,9 +8,10 @@ import gun_interface
 
 const
   PowerBins*    = [1.0, 1.5, 2.0, 3.0]  ## 4 bins; ponytail: fixed array, add runtime config if needed
-  WindowSize*   = 50                      ## rolling window ticks for fitness
+  WindowSize*   = 100                     ## rolling window ticks for fitness
   MaxBullets*   = 2048                    ## hard cap; ponytail: ring buffer, resize if more guns added
-  MinHitRate*   = 0.40                    ## 40% threshold for acceptable power selection
+  MinHitRate*          = 0.40   ## 40% threshold for acceptable power selection
+  MinObsBeforeCompete* = 15    ## min observations before a gun×bin enters competition
 
 type
   GunId*   = int  ## index into the guns seq
@@ -124,12 +125,30 @@ proc bestPower*(t: VirtualTracker, gunId: GunId): (int, float) =
 
 proc bestGun*(t: VirtualTracker): GunId =
   ## Pick gun with highest hit rate across all power bins.
+  ## Guns with fewer than MinObsBeforeCompete observations across all bins are
+  ## skipped unless every gun is below threshold (then fall back to best of all).
   ## ponytail: O(n*bins), fine for small gun counts
-  var bestRate = -1.0
-  result = 0
+  proc bestAmong(t: VirtualTracker, requireMin: bool): GunId =
+    var bestRate = -1.0
+    result = 0
+    for gunId in 0..<t.fitness.len:
+      var maxCount = 0
+      for binIdx in 0..<len(PowerBins):
+        maxCount = max(maxCount, t.fitness[gunId].bins[binIdx].count)
+      if requireMin and maxCount < MinObsBeforeCompete: continue
+      for binIdx in 0..<len(PowerBins):
+        let r = t.fitness[gunId].bins[binIdx].hitRate()
+        if r > bestRate:
+          bestRate = r
+          result = gunId
+
+  # Try gated selection first; fall back to ungated if nothing qualifies
+  var anyQualifies = false
   for gunId in 0..<t.fitness.len:
     for binIdx in 0..<len(PowerBins):
-      let r = t.fitness[gunId].bins[binIdx].hitRate()
-      if r > bestRate:
-        bestRate = r
-        result = gunId
+      if t.fitness[gunId].bins[binIdx].count >= MinObsBeforeCompete:
+        anyQualifies = true
+        break
+    if anyQualifies: break
+
+  result = if anyQualifies: t.bestAmong(true) else: t.bestAmong(false)