"""Independent round8 source/optimizer/comparison audits; no old reports rewritten."""
from pathlib import Path
import sys
import types

PROJECT=Path(__file__).resolve().parents[1];sys.path.insert(0,str(PROJECT));sys.path.insert(0,str(PROJECT/'scripts'))
from pallet.cluster_targets import digest
from pallet.mask_model_training_data import read,write,require
from pallet.pallet_description_round8_data import ROOT,sources_current

path=PROJECT/'scripts/finalize_pallet_description_round7.py'
SOURCE=path.read_text().replace('round7','round8')
for old,new in (('validate_required,POLICY,REINSPECTION','validate_required,POLICY'),
                ('read(POLICY),read(REINSPECTION)','read(POLICY)')):
    require(SOURCE.count(old)==1,'Round8 source-audit anchor changed');SOURCE=SOURCE.replace(old,new)
implementation=types.ModuleType('_round8_audits');implementation.__file__=__file__
exec(compile(SOURCE,str(path),'exec'),implementation.__dict__)
source_replay=implementation.source_replay;training=implementation.training;bundles=implementation.bundles;paired_metrics=implementation.paired_metrics


def report():
    require(read(ROOT/'training_verification.json')['state']=='round8_actual_training_and_latest_source_replay_passed','Actual training audit required')
    paired_metrics();data=read(ROOT/'preparation_report.json');replay=read(ROOT/'source_replay.json');mask=read(ROOT/'mask_regression_audit.json');vlm=read(ROOT/'vlm_regression_audit.json')
    sources=data['sources'];prior=replay['retained_optional_prior_baselines']
    lines=['# 第八轮真实训练与同标签比较','',f"冻结{data['targets']}目标／{data['distinct_images']}张不同原图，40细类／9大类；新增{data['newly_added_targets']}，更新{data['updated_existing_targets']}已有目标，{prior}项非必做诊断保留原接受基线。训练{data['splits']['train']}、组件留出{data['splits']['holdout']}。",'',
        f"第七轮来源：新图{sources['new']['saved']}项correct／{sources['new']['events']}事件，已有{sources['existing']['saved']}项correct／{sources['existing']['events']}事件。优先／固定抽检待办为零；非必做诊断不算本轮完成，不接受草稿／暂缓，不制造correct事件。",'',
        '## 同一份最新留出标签','', '| 指标 | 第七轮独立分支 | 第八轮独立分支 |','|---|---:|---:|',
        f"| 细类Top1 | {mask['old_fine_top1']:.2%} | {mask['new_fine_top1']:.2%} |",f"| 大类Top1 | {mask['old_parent_top1']:.2%} | {mask['new_parent_top1']:.2%} |",f"| IoU中位数 | {mask['old_iou_median']:.4f} | {mask['new_iou_median']:.4f} |",'',
        f"VLM完整JSON：{vlm['metrics']['old']['strict_complete_JSON_valid']}/{vlm['metrics']['old']['targets']} → {vlm['metrics']['new']['strict_complete_JSON_valid']}/{vlm['metrics']['new']['targets']}。结构字段一致率含明确结构恢复："]
    for f in ('支腿','横梁','材质'):
        a=vlm['attribute_recovery']['old']['attribute_exact_including_explicit_structure_recovery'][f]['accuracy'];b=vlm['attribute_recovery']['new']['attribute_exact_including_explicit_structure_recovery'][f]['accuracy'];lines.append(f'- {f}：{a:.2%} → {b:.2%}。')
    lines+=['','完整逐类／逐图改善和退步均保留，不宣称全面提升，不按留出指标挑检查点。分类／轮廓独立分支从固定通用ResNet18起训24轮，评估后全量20轮；VLM独立分支从固定Qwen3.5-0.8B基座起训2轮，全量只续训第七轮全量适配器2轮。全部优化器、数据暴露、权重及缺失观察引号／EOS监督已核验。',
        f"结构字段冲突{data['semantic_conflict_targets']}项；不按模板重写标签，仅屏蔽冲突字段监督。缺少可靠观察不制造正常GT，模型非空观察及合法JSON不等于语义准确率认证。",'',
        '组件／同图划分继承，反复迭代的组件留出不是新采集盲测。旧模型、冻结数据、正式标签、库、历史、网址及草稿namespace保持。不恢复文字，不自动晋升GT／机器人资格，不提交／推送。',
        '本报告只证明真实训练及比较；下一轮推理和任务须单独验收，不能把训练完成当作已发布。']
    require(sources_current(),'Source changed before report');(ROOT/'TRAINING_REPORT.md').write_text('\n'.join(lines)+'\n')
    write(ROOT/'TRAINING_HANDOFF.json',dict(state='round8_actual_training_source_replay_and_matched_comparison_complete',data=data,mask=mask,vlm=vlm,source_replay_sha256=digest(ROOT/'source_replay.json'),
        training_verification_sha256=digest(ROOT/'training_verification.json'),report_sha256=digest(ROOT/'TRAINING_REPORT.md'),old_models_data_and_histories_preserved=True,production_annotation_POSTs=0,no_commit_or_push=True))
    print(dict(state='round8_training_audited',old_fine=mask['old_fine_top1'],new_fine=mask['new_fine_top1']))


def deliver():
    import json
    from urllib.request import urlopen
    from pallet.pallet_description_round8_release import REVIEW_SETS,RELEASE
    from pallet.pallet_review_ptunnel import profiles
    from pallet.mask_class_note_integration import snapshot_db
    from reload_mask_taxonomy_keyboard_ui import state
    from serve_structure_review import process_identity
    from serve_pallet_description_round8 import configure
    require(read(ROOT/'TRAINING_HANDOFF.json')['state']=='round8_actual_training_source_replay_and_matched_comparison_complete','Training report required')
    prior=read(ROOT/'before_round8_publication.json');routes=profiles();services={};proofs={}
    require(all(all(routes[k].get(f)==v.get(f) for f in ('id','host','port','target','url','tunnelID')) for k,v in prior['saved_ptunnel_profiles'].items()),'Old ptunnel profile changed')
    require(all(digest(PROJECT/n)==h for n,h in prior['protected_training'].items()),'Old training dataset/model changed')
    for name,v in prior['old_services'].items():
        current=read(PROJECT/f'outputs/pallet_description_{name}_managed_v1/supervisor.json')
        require(current['pid']==v['pid'] and current['app']['pid']==v['app']['pid'] and current['public_url']==v['public_url'] and current['process_identity']==process_identity(current['pid']),'Old app/supervisor/URL changed')
    for mode,s in REVIEW_SETS.items():
        configure(mode);folder=RELEASE/mode;current=read(s['managed']/'supervisor.json');proof=read(s['managed']/'round8_public_verification.json');pointer=read(folder/'pointer_verification.json')
        require(current['state']=='connected' and current['process_identity']==process_identity(current['pid']),'Round8 supervisor unhealthy')
        require(proof['state']=='round8_ptunnel_authenticated_public_readonly_passed' and not proof['page_errors'] and not proof['blocked_writes'] and proof['production_annotation_POSTs']==proof['production_decision_POSTs']==0,'Public read-only acceptance required')
        require(proof['implementation_sha256']==digest(PROJECT/'scripts/verify_pallet_description_round8_public.py'),'Public checker changed')
        require(pointer['state']=='right_pan_left_edit_shortcut_guards_passed' and len(pointer['viewports'])==2 and not pointer['page_errors'],'Actual pointer/keyboard guard checks required')
        snapshot_db(s['reviews']/'mask_reviews.sqlite3',folder/'final.sqlite3');bm,be,_=state(folder/'before_public.sqlite3');am,ae,ar=state(folder/'final.sqlite3')
        require(bm==am and ae[:len(be)]==be and ar=={k:(rev,raw) for _,k,rev,raw in ae},'Production history/current row preservation failed')
        require(all('SYNTHETIC' not in raw for _,k,rev,raw in ae),'Synthetic production annotation found')
        services[mode]=dict(public_url=current['public_url'],targets=read(s['root']/'manifest.json')['request_count'],namespace=s['namespace'],events=len(ae))
        proofs[mode]=dict(browser_sha256=digest(folder/'browser_verification.json'),pointer_sha256=digest(folder/'pointer_verification.json'),public_sha256=digest(s['managed']/'round8_public_verification.json'))
    health={}
    for port in range(3413,3426):
        with urlopen(f'http://127.0.0.1:{port}/healthz',timeout=10) as response:h=json.load(response)
        require(h['status']=='ok','Old/new app unhealthy');health[str(port)]=h['publication_version']
    require(sources_current(),'Frozen sources changed before delivery');selection=read(ROOT/'release_selection.json');priority=read(ROOT/'review_priority_selection.json')
    lines=[(ROOT/'TRAINING_REPORT.md').read_text(),'','## 第八轮人工复核入口已验收','',
        f"- 新图类别／轮廓／语义：{services['new']['targets']}项，{services['new']['public_url']} 。",
        f"- 已有重点复核：{priority['required']}项（优先{priority['tier_counts'].get('priority',0)}＋固定{priority['tier_counts'].get('audit',0)}抽检），{services['existing']['public_url']} 。其余{priority['tier_counts'].get('diagnostic',0)}项诊断非必做；全部{services['existing']['targets']}项和历史仍可查。",'',
        '沿用原访问码，新域名需登录。独立3424／3425、任务／库／草稿namespace及ptunnel profiles，所有旧入口和进度保持，不搬运或删除旧草稿，已完成的Class13临时组不重新开启。',
        f"每类100张新图总目标4,000，本次严格K、定位及跨轮原图去重通过{services['new']['targets']}张，缺口{selection['gap']}；不补点、不放宽阈值、不复制模板、不混入旧图。",'',
        '## 验收与边界','',
        '- 415项CPU回归通过；新图与已有任务分别通过桌面／390宽临时库快捷键、严格K、409、草稿／备份、存储失败、忙锁、D／A固定序列、所有批量预览、替换精确分项／去重／幂等／整批409与全队列重放；另有左右键实测和拖动期间快捷键不提交。',
        '- 已有队列验证优先／固定抽检、刷新不补入新抽检、独立疑点结论幂等与409、40细类＋全部选项的筛选矩阵。完成记录可查，批量范围在打开详情时固定。',
        '- Ctrl+S／⌘+S保存为已复核且停留，D完成后下一条，A不提交；默认复核人1、空异常默认正常、右键平移／左键修点、历史追加、草稿备份保留。描述增加／删除／替换只按中文逗号完整项精确匹配，不改其他独有描述。',
        '- 公网只登录、看图、读取与预览；生产模拟标注／疑点结论POST均为零。旧模型、冻结数据、正式标签、历史、服务PID和ptunnel profiles保持，不自动晋升GT／机器人资格，未提交／推送。',
        '- Linux Chromium不代表真实MacOS／iPhone Safari实机验收；未配置开机自启，不承诺永久在线或永久域名。',
        '完整证据：本目录TRAINING_HANDOFF.json、training_verification.json、mask_regression_audit.json、vlm_regression_audit.json，以及outputs/pallet_description_simplified_v1/round8_v1/。']
    (ROOT/'REPORT.md').write_text('\n'.join(lines)+'\n')
    write(ROOT/'HANDOFF.json',dict(state='round8_real_training_and_bounded_next_review_delivered_ptunnel',data=read(ROOT/'preparation_report.json'),services=services,priority=priority,quota_gap=selection['gap'],proofs=proofs,
        training_handoff_sha256=digest(ROOT/'TRAINING_HANDOFF.json'),training_verification_sha256=digest(ROOT/'training_verification.json'),bundle_source_replay_sha256=digest(ROOT/'bundle_source_replay.json'),report_sha256=digest(ROOT/'REPORT.md'),
        healthy_ports=health,regression_tests=415,old_models_data_histories_services_profiles_preserved=True,production_annotation_POSTs=0,production_decision_POSTs=0,no_commit_or_push=True))
    print(dict(state='round8_delivered',services=services,priority=priority['tier_counts'],production_annotation_POSTs=0))


if __name__=='__main__':
    import argparse
    p=argparse.ArgumentParser();p.add_argument('phase',choices=('source','training','bundles','report','deliver'));a=p.parse_args();{'source':source_replay,'training':training,'bundles':bundles,'report':report,'deliver':deliver}[a.phase]()
