"""Count large fixtures with the official *reference*, not proven native tokenizer.

An offline study only. No outputs here may be called exact native token counts.
"""
import argparse
import json
import time
from pathlib import Path
from context_adapter import DecisionContext, adapt, canonical
from decision_policy import eligible_tools, intent_eligible, retrieve
from scripts.token_accounting_study import reference, fast_bpe
from scripts.final_qualification import cases


def main():
    p=argparse.ArgumentParser();p.add_argument('--wheel',required=True);p.add_argument('--assets',required=True)
    a=p.parse_args();ns=reference(a.wheel)
    meta=ns['parse_tokenizer_blob'](ns['read_tokenizer_blob'](str(Path(a.assets)/'needle3.cact')))
    tok=ns['RefTokenizer'](meta);tok._bpe=lambda text:fast_bpe(tok,text)
    rows=[]
    # One placement per size/family for tokenizer cost. HTTP evaluation covers all three placements.
    for case in cases():
        if case['group']!='long-context' or not case['id'].endswith('-middle'):continue
        b=case['request'];text=b['messages'][1]['content'];start=time.perf_counter()
        count=len(tok.encode(text));ms=(time.perf_counter()-start)*1000
        policy=DecisionContext(**b['decision_context']);reduced=adapt(b['messages'],policy,b['tools'])
        compact=None;compact_bytes=None
        if reduced.sufficient:
            tools,_=eligible_tools(b['tools'],policy)
            tools=intent_eligible(reduced.inference_text(),b['tools'],tools)
            tools,_=retrieve(reduced.inference_text(),tools)
            bindings={t['name']:policy.arguments[t['name']] for t in tools if t['name'] in policy.arguments}
            system=reduced.inference_system()
            if bindings:system+='\nCurrent exact argument bindings: '+canonical(bindings)
            visible=system+reduced.inference_text()+canonical(tools)
            compact=len(tok.encode(visible));compact_bytes=len(visible.encode())
        rows.append({'id':case['id'],'tool_output_utf8_bytes':len(text.encode()),'reference_tool_output_tokens':count,'reference_tokenization_ms':ms,'inference_sufficient':reduced.sufficient,'reference_visible_compact_tokens':compact,'visible_compact_utf8_bytes':compact_bytes,'native_hidden_template_tokens':None,'native_query_tokens':None})
    print(json.dumps({'count_kind':'official_newer_reference_not_native_verified','native_exact_accounting':False,'rows':rows,'limitations':['Counts apply to tool-output text and visible compact concatenation; no native hidden template count implied.','Size labels still denote records. Actual reference token counts are reported independently.','Standalone reference encoder timing is not pinned native tokenizer latency.']},indent=2))

if __name__=='__main__':main()