"""Is the earlier bit-identity failure reproducible? Stress the pattern that failed:
  warm W  -> P+ (REUSED) -> P+ (COLD),  repeated N times, and count mismatches.
If mismatches recur, it is a race/contamination bug. If not, the earlier failure had another cause.
"""
import json, urllib.request, hashlib
URL="http://127.0.0.1:8300/v1/chat/completions"
def ask(prompt, mt=8, temp=0):
    body={"model":"glm-5.3-exl3","messages":[{"role":"user","content":prompt}],"max_tokens":mt,
          "temperature":temp,"return_token_ids":True}
    r=urllib.request.Request(URL,data=json.dumps(body).encode(),headers={"Content-Type":"application/json"})
    d=json.loads(urllib.request.urlopen(r,timeout=3600).read())
    st=d.get("tensorfold",{}) or {}
    ids=st.get("token_ids") or []
    return st.get("cached"), hashlib.sha256(json.dumps(ids).encode()).hexdigest()[:16]
unit="The service validated the payload and emitted a structured record with stable ids. "
bad=0; tot=0
for it in range(8):
    base=(unit*(60+it*7))                      # a DIFFERENT base each iteration (new conversation)
    W=base+"\nReply with YES."
    P=base+"\nReply with OK. Be brief."        # P+ pattern: multi-token suffix
    cW,sW=ask(W)
    cR,sR=ask(P)
    cC,sC=ask(P)
    ok = (sR==sC)
    tot+=1; bad += (not ok)
    print(f"  it{it}: warm_cached={cW:5} reused_cached={cR:5} cold_cached={cC:5} "
          f"reused={sR} cold={sC} MATCH={ok}", flush=True)
print(f"\n  mismatches: {bad}/{tot}")
print("### STRESS DONE")