"""Replay real new labels/training, report the matched comparison, retain old delivery."""
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,jsonl,require
from pallet.pallet_description_round6_data import ROOT,sources_current
from pallet.pallet_description_round6_release import REVIEW_SETS,RELEASE

path=PROJECT/'scripts/finalize_pallet_description_round5.py';SOURCE=path.read_text().replace('round5','round6')
implementation=types.ModuleType('_round6_final_audit');implementation.__file__=__file__
exec(compile(SOURCE,str(path),'exec'),implementation.__dict__)
source_replay=implementation.source_replay;training=implementation.training;bundles=implementation.bundles


def metrics():
    data=read(ROOT/'preparation_report.json');mask=read(ROOT/'mask_run_v1/independent_branch/holdout_metrics.json');oldmask=read(ROOT/'mask_run_v1/old_component_branch_comparison.json')
    vlm=read(ROOT/'vlm_run_v1/component_branch/holdout_metrics.json');oldvlm=read(ROOT/'vlm_run_v1/old_component_comparison/holdout_metrics.json')
    recovery=read(ROOT/'vlm_run_v1/component_branch/structure_recovery_audit.json');oldrecovery=read(ROOT/'vlm_run_v1/old_component_comparison/structure_recovery_audit.json')
    semantic={r['target_key']:r for r in jsonl(ROOT/'description_data/records.jsonl')};predictions=jsonl(ROOT/'vlm_run_v1/component_branch/holdout_predictions.jsonl')
    positive=[p for p in predictions if semantic[p['target_key']]['observation_supervised']]
    answered=sum(p['valid'] and bool(p['fields'].get('观察')) for p in positive)
    return data,mask,oldmask,vlm,oldvlm,recovery,oldrecovery,positive,answered


def report():
    require(read(ROOT/'training_verification.json')['state']=='round6_actual_training_and_latest_source_replay_passed','Actual training audit required')
    require(read(ROOT/'bundle_source_replay.json')['state']=='round6_next_tasks_latest_human_baseline_and_model_raw_replayed','Task source audit required')
    data,mask,oldmask,vlm,oldvlm,recovery,oldrecovery,positive,answered=metrics();access={};counts={}
    for mode,s in REVIEW_SETS.items():
        require(read(RELEASE/mode/'round6_browser_verification.json')['state']=='passed_round6_actual_task_desktop390_all_batch_and_fast_navigation','Browser safety audit required')
        public=read(s['managed']/'round6_public_verification.json');require(public['state']=='round6_ptunnel_authenticated_public_readonly_passed','ptunnel read-only audit required')
        access[mode]=public['public_url'];counts[mode]=read(s['root']/'manifest.json')['request_count']
    lines=['# 第六轮训练与下一轮人工复核','',f"冻结数据：{data['targets']}目标／{data['distinct_images']}张不同原图／40细类／9大类。新增{data['newly_added_targets']}个目标，更新{data['updated_existing_targets']}个已有目标；不重复计数。训练{data['splits']['train']}、组件留出{data['splits']['holdout']}。",'',
        '来源为第五轮真实完成的432新图和1,098已有复核；1,330／1,683条历史事件原样保留。旧训练集、权重、任务和ptunnel入口未替换。','',
        '## 同一份最新标签组件留出对比','', '| 指标 | 第五轮独立模型 | 第六轮独立模型 |','|---|---:|---:|',
        f"| 细类Top1 | {oldmask['fine_top1']:.2%} | {mask['fine_top1']:.2%} |",f"| 大类Top1 | {oldmask['parent_top1']:.2%} | {mask['parent_top1']:.2%} |",f"| mask IoU中位数 | {oldmask['segmentation']['iou_median']:.4f} | {mask['segmentation']['iou_median']:.4f} |",'',
        '本轮独立细分类回退，不宣称整体提升，也不自动替换正式模型。新、旧权重均保留；逐类与逐目标配对回退见mask_regression_audit.json。后续全量模型输出仅作为需要人工复核的草稿，不以训练集拟合成绩抵消独立留出回退。','',
        f"独立分支重新从固定通用ResNet18预训练起训，24轮；随后全量辅助分支20轮。训练侧缺少细类{mask['classes_absent_from_training']}，不能声称这些类已独立验证。保留原图／采集关联分量划分，不按新标签重分配。旧独立模型未训练本次留出目标，但留出反复迭代、样本曾被全量模型筛选，不是新采集前瞻盲测。",'',
        '## VLM','',f"Qwen3.5-0.8B，冻结视觉、rank8 LoRA；独立分支从固定基座训练2轮，全量分支只续训第五轮全量适配器2轮。保留相同提示词、256 token预算和配置。完整JSON：旧{oldvlm['strict_complete_JSON_valid']}/{oldvlm['targets']} → 新{vlm['strict_complete_JSON_valid']}/{vlm['targets']}。",'',
        f"结构字段冲突目标{data['semantic_conflict_targets']}项，冲突字段遮蔽而非模板改写。观察监督{data['semantic_supervised_fields']['观察']}项；缺少可靠观察时仅监督空字符串引号／JSON／EOS，不构造正常语义负例。完整JSON有效不等于内容正确。"]
    lines += ['本轮VLM材质、支腿和JSON完整率出现回退，横梁略升，不宣称整体语义提升。保留旧适配器及逐条原始输出；不通过增加非空观察数量来冒充观察准确率。逐类／逐项对比见vlm_regression_audit.json。']
    for k in ('支腿','横梁','材质'):lines.append(f"- {k}一致率（包含明确结构恢复，分母为全部有该字段参考的留出项）：{oldrecovery['attribute_exact_including_explicit_structure_recovery'][k]['accuracy']:.2%} → {recovery['attribute_exact_including_explicit_structure_recovery'][k]['accuracy']:.2%}。")
    lines += [f"- {len(positive)}项有人工观察参考的留出图中，{answered}项输出有效非空观察，{len(positive)-answered}项弃答／无效；非空不代表正确，异常泛化未认证。",'',
        '## 下一轮复核入口','']
    for mode,title in (('new','新图类别／轮廓／语义'),('existing','已有定向语义与联合复核')):lines.append(f"- **{title}：{counts[mode]}项**，{access[mode]}，沿用原访问码。")
    selection=read(ROOT/'release_selection.json');lines += [f"- 每细类100张新原图的总目标4,000，严格通过且跨轮去重的新图{counts['new']}，缺口{selection['gap']}；未混入旧图、补点、复制模板或放宽门槛。逐类明细见release_quota_report.json。",'- 已有队列保留最新人工类别、轮廓和描述；新版模型原文只读对照，分歧仅为复核线索，不自动判人工错误。']
    for mode in ('new','semantic'):
        d=read(ROOT/f'description_inference_{mode}.json');lines.append(f"- {mode}全量模型推理：JSON {d['schema_valid']}/{d['targets']}，无效观察拒绝{d['observation_withheld']}，有效空观察／未知{d['strict_valid_empty_observation']}。")
    lines += ['','## 验收和边界','', '- 第六轮使用独立3420／3421、任务／库／草稿namespace和独立ptunnel profiles；不替换第五轮3418／3419及历史服务。只用ptunnel，不回退trycloudflare。',
        '- 保留默认复核人1、D成功保存后下一条、A不提交、右键平移／左键修点、严格K、409、历史恢复、本机草稿、原子保存、覆盖／结构选项／描述项增加和删除全部批量功能。',
        '- 保留完整来源字节并行SHA-256和图像预取；人工revision实时读取，不用TTL／抽样。CPU回归、临时库桌面／390宽及全序列重放证据保留。',
        '- 公网只登录、看图和只读预览／回查，未在生产提交模拟标注；未测试真实MacOS／iPhone Safari实机。',
        '- 空模型观察表示未知，应用层仍按用户规则空异常默认正常。没有改正式GT、自动晋升训练／机器人资格、覆盖旧模型或提交／推送。']
    require(sources_current(),'Source changed before delivery');(ROOT/'REPORT.md').write_text('\n'.join(lines)+'\n')
    write(ROOT/'HANDOFF.json',dict(state='round6_real_training_and_next_review_delivered_ptunnel',datasets=data,mask_holdout=mask,old_mask=oldmask,vlm_holdout=vlm,old_vlm=oldvlm,
        review_counts=counts,access=access,full_mask_sha256=read(ROOT/'mask_run_v1/training_complete.json')['full_pool'],vlm_adapter_sha256=read(ROOT/'vlm_run_v1/training_complete.json')['adapter_sha256'],
        report_sha256=digest(ROOT/'REPORT.md'),production_annotation_POSTs=0,old_models_data_and_histories_preserved=True,no_commit_or_push=True))


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