#!/usr/bin/env python3
"""Dated post-hoc replay of saved essai1 trees; no refitting or source changes."""
import argparse, hashlib, json
from pathlib import Path
ROOT=Path(__file__).resolve().parent
SOURCE=ROOT.parent/'public'

def sha(p):return hashlib.sha256(p.read_bytes()).hexdigest()
def calculate():
 ns={'__file__':str(SOURCE/'experiment.py'),'__name__':'posthoc_replay'}
 exec(compile((SOURCE/'experiment.py').read_text(),str(SOURCE/'experiment.py'),'exec'),ns)
 ns['verify']()
 data=json.loads((SOURCE/'data.json').read_text());original=json.loads((SOURCE/'results.json').read_text());cases={c['id']:c for c in data['cases']}
 rows=[]
 for split in data['splits']:
  saved=next(f for f in original['folds'] if f['fold']==split['fold'])
  for use,key in [('explore','test_explore'),('decide','test_decide')]:
   for cid in split[key]:
    c=cases[cid];before,_,_,_=ns['predict_tree'](saved['trees']['initial'],c['x']);after,path,support,fallback=ns['predict_tree'](saved['trees']['corrected'],c['x'])
    rows.append({'case':cid,'fold':split['fold'],'use':use,'features':c['x'],'truth':c['truth'],'before':before,'after':after,'corrected_path':path,'leaf_support':support,'fallback':fallback})
 def metrics(group,column):return ns['metrics']([{'truth':r['truth'],'prediction':r[column]} for r in group])
 by_use={use:{'before':metrics([r for r in rows if r['use']==use],'before'),'after':metrics([r for r in rows if r['use']==use],'after')} for use in ['explore','decide']}
 # Replay must reproduce all three already published quantities exactly.
 assert by_use['explore']['before']==original['aggregate']['same_use']['learner']
 assert by_use['decide']['before']==original['aggregate']['changed_use']['learner']
 assert by_use['decide']['after']==original['aggregate']['after_correction']['learner']
 return {'id':'PW238-essai01-posthoc-retention-20261019','date':'2026-10-19','status':'post-hoc diagnostic; no new training; not in the original frozen protocol',
 'source_sha256':{n:sha(SOURCE/n) for n in ['experiment.py','data.json','results.json','protocol.json','freeze.json']},
 'scope':'Replay both saved initial and corrected trees on both existing held-out uses; all substantive-bundle splits retained.',
 'combined_weighting':'Equal weight to the 96 authored exploratory and 96 authored decision cases; not an estimated deployment mixture.',
 'by_use':by_use,'combined':{'before':metrics(rows,'before'),'after':metrics(rows,'after')},
 'original_use_regressions':[r for r in rows if r['use']=='explore' and r['before']==r['truth'] and r['after']!=r['truth']],
 'predictions':rows}

if __name__=='__main__':
 p=argparse.ArgumentParser();p.add_argument('action',choices=['write','verify']);a=p.parse_args();result=calculate();target=ROOT/'results.json'
 if a.action=='write':
  assert not target.exists(),'Refusing to overwrite dated finding'
  target.write_text(json.dumps(result,indent=2)+'\n')
 else:assert json.loads(target.read_text())==result,'Post-hoc replay differs from saved finding'
 print(json.dumps({'by_use':{u:{k:v['correct'] for k,v in m.items()} for u,m in result['by_use'].items()},'combined':{k:v['correct'] for k,v in result['combined'].items()},'regressions':[r['case'] for r in result['original_use_regressions']]},indent=2))
