## diag_ppo_fullbuffer.nim — run ppoUpdate on a FULL 8192-transition buffer.
## Verifies whether the round-10 death (first ppoUpdate at buffer capacity)
## is a real crash in ppoUpdate or purely the saveAdamStates empty-shape bug.

import std/[math, random]
import arraymancer
import PPO_Bot/network
import PPO_Bot/training

var ac = initActorCritic()
var adam: ACAdamStates  # uninitialised → ppoUpdate must reinit (fresh-process path)
var buf = initTrajectoryBuffer()

randomize(1)
while buf.len < MAX_TRANSITIONS:
  var t: Transition
  for i in 0..<STATE_DIM: t.state[i] = rand(1.0'f32) - 0.5'f32
  for i in 0..<ACTION_DIM: t.action[i] = rand(1.0'f32) - 0.5'f32
  t.logProb = rand(1.0'f32) - 1.0'f32
  t.reward = rand(0.02'f32) - 0.01'f32
  t.value = rand(0.1'f32)
  t.done = buf.len mod 300 == 299
  buf.add(t)

echo "buffer.len = ", buf.len, " (MAX=", MAX_TRANSITIONS, ")"
var ac2 = ac
var adam2: ACAdamStates
try:
  let m = ppoUpdate(ac2, buf, lastValue = 0.0'f32, adamStates = adam2)
  echo "ppoUpdate OK: aLoss=", m.actorLoss, " vLoss=", m.valueLoss, " gNorm=", m.gradNorm
except CatchableError as e:
  echo "CAUGHT CatchableError: ", e.msg
  echo getStackTrace(e)
except Defect as e:
  echo "CAUGHT Defect: ", e.msg
  echo getStackTrace(e)