"""Read-only training audit/report after real workers finish; no fabricated completion."""
from collections import Counter
from datetime import datetime,timezone
import json
from pathlib import Path
import sys
import xml.etree.ElementTree as ET

PROJECT=Path(__file__).resolve().parents[1];sys.path.insert(0,str(PROJECT))
from pallet.mask_model_training_data import read,write,jsonl,require,sha
from pallet.cluster_targets import digest
from pallet.pallet_description_round4_data import ROOT,verify,sources_current
from pallet.pallet_description_round4_release import REVIEW_SETS,RELEASE


def audit_training():
    import numpy as np
    import cv2
    from pallet.pose_review_geometry import raster_mask
    from pallet.mask_template_refined import readonly
    from contextlib import closing
    manifest,records=verify();require(sources_current(),'Live sources changed; do not claim latest publication')
    masks=ROOT/'mask_run_v1';vlm=ROOT/'vlm_run_v1';require((masks/'training_complete.json').exists() and (vlm/'training_complete.json').exists(),'Both actual trainings must finish')
    originals={}
    for mode in ('new','existing'):
        props={r['target_key']:r for r in jsonl(PROJECT/f'outputs/pallet_description_round3_{mode}_v1/proposals.jsonl')}
        with closing(readonly(ROOT/f'private/snapshots/{mode}.sqlite3')) as db:saved={k:json.loads(raw) for k,raw in db.execute('SELECT target_key,payload FROM reviews')}
        originals.update({k:saved.get(k,p) for k,p in props.items()})
    for r in records:
        old=originals[r['target_key']];require(r['polygon']==old['polygon'] and r['class_id']==(148 if old['class_id']==149 else old['class_id']),'Latest human labels do not replay')
        full=raster_mask(r['polygon'],*r['source']['size']);x1,y1,x2,y2=r['roi'];mask=cv2.imread(str(ROOT/r['mask']),0)>0;valid=cv2.imread(str(ROOT/r['valid']),0)>0
        require(np.array_equal(mask,full[y1:y2,x1:x2]) and mask.sum()==full.sum() and valid[mask].all(),'Native mask clipped or invalid')
    steps={}
    for phase in ('independent_branch','full_pool'):
        folder=masks/phase;receipt=read(folder/'training_receipt.json');history=[read(p) for p in sorted(folder.glob('epoch_*.json'))]
        require(receipt['optimization_steps']==sum(v['epoch_updates'] for v in history)>0 and receipt['checkpoint_sha256']==digest(folder/'final_ema.pt'),'Mask optimizer/checkpoint evidence missing')
        require(all(np.isfinite(list(v['mean_losses'].values())).all() for v in history),'Nonfinite mask history');steps['mask_'+phase]=receipt['optimization_steps']
    for phase in ('component_branch','full_pool'):
        folder=vlm/phase;receipt=read(folder/'training_receipt.json');plan=read(folder/'execution_plan.json');audit=read(folder/'token_alignment_audit.json')
        require(receipt['optimizer_steps']>0 and receipt['adapter_sha256']==digest(folder/'adapter/adapter_model.safetensors'),'VLM optimizer/checkpoint evidence missing')
        require(all(v['eos_supervised'] for v in audit) and plan['config']['missing_or_conflicting_field_value_loss_masked'],'VLM mask/EOS evidence missing')
        require(plan['independent_base_start']==(phase=='component_branch'),'Independent VLM warm-start contaminated')
        steps['vlm_'+phase]=receipt['optimizer_steps']
    semantics=jsonl(ROOT/'description_data/records.jsonl');require(all(not r['missing_observation_is_normal'] for r in semantics),'Missing observations promoted to normal')
    write(ROOT/'training_verification.json',dict(state='passed_latest_snapshot_mask_and_partial_semantic_training_replay',targets=len(records),
        optimizer_steps=steps,dataset_manifest_sha256=digest(ROOT/'dataset_manifest.json'),latest_source_labels_replayed=True,no_fabricated_correct_events=True,
        independent_branches_checked=True,observation_format_not_semantic_certification=True,verified_at=datetime.now(timezone.utc).isoformat()))


def report():
    require((ROOT/'training_verification.json').exists(),'Actual training replay audit required')
    require(read(ROOT/'bundle_source_replay.json')['state']=='both_next_round_bundles_independently_replayed','Independent baseline/draft replay required')
    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_structure_recovery_audit.json')['models'];newrec=recovery['component_branch'];oldrec=recovery['old_component_comparison']
    q=read(ROOT/'release_selection.json');counts={mode:read(s['root']/'manifest.json')['request_count'] for mode,s in REVIEW_SETS.items()}
    access={mode:read(s['managed']/'public_access.json') for mode,s in REVIEW_SETS.items()}
    for mode,s in REVIEW_SETS.items():
        require(read(RELEASE/mode/'round4_browser_verification.json')['state']=='passed_round4_actual_task_desktop_and390_safety','Browser acceptance missing')
        require(access[mode].get('tunnel_connected') and access[mode].get('public_url'),'Public service not connected')
        public=read(s['managed']/'public_verification.json')
        require(public['state']=='passed_authenticated_round4_public_assets_catalog_and_per_class_targets' and public['manifest_sha256']==digest(s['root']/'manifest.json'),'Authenticated public readback required')
    lines=['# 第四轮训练与下一轮复核交付','',f"数据：{data['targets']}目标／{data['distinct_images']}张不同原图／40细类／9大类；训练{data['splits']['train']}、组件留出{data['splits']['holdout']}。149→148已纳入，原草稿／完成动作及历史均未改写。",'',
        '## 分类／轮廓：同一最新标签组件留出集','', '| 指标 | 旧独立分支 | 新独立分支 |','|---|---:|---:|',
        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：旧{oldmask['segmentation']['iou_median']:.4f}，新{mask['segmentation']['iou_median']:.4f}。新分支训练侧缺少这些细类样本：{mask['classes_absent_from_training']}，相关细类留出成绩不能当作已学会该类的证据；不为提高指标按标签重分留出。全量标注模型随后覆盖全部40类。",
        '本次新模型的分类留出成绩没有提升，不自动认定为更优生产模型；仅作为已更新标签的下一轮人工标注辅助，旧模型继续保留。','',
        '旧模型149与148的概率合并后按最新40类标签比较；两独立分支均未训练留出分量。部分图由此前全量模型筛选，因此仍不是新采集的前瞻盲测。训练拟合成绩不当作独立成绩。','',
        '## VLM：完整JSON与部分人工语义','',f"- 旧分支完整JSON：{oldvlm['strict_complete_JSON_valid']}/{oldvlm['targets']}；新分支：{vlm['strict_complete_JSON_valid']}/{vlm['targets']}。",
        '- 两者使用各自冻结提示词、同一当前留出图像与标签；生成预算均为256。旧分支只训练过旧平衡子集，新分支使用新版全部有监督语义训练目标，不能解读为只改变单一因素的消融。',
        f"- 新分支除{vlm['strict_complete_JSON_valid']}项完整JSON外，还有{newrec['format_counts']['structure_only_recovered']}项只恢复结构，观察不采用；结构不可恢复{newrec['format_counts']['no_structure_available']}项。这不计为完整JSON成功。",
        '- 只按实际合法结构输出（含显式结构恢复），以全部有对应标签的留出项为分母：',
        *[f"  - {k}：旧{oldrec['attribute_structure_recovery'][k]['accuracy_all_labelled']:.2%}，新{newrec['attribute_structure_recovery'][k]['accuracy_all_labelled']:.2%}。" for k in ('支腿','横梁','材质')],
        f"- 当前冻结包含{data['semantic_supervised_fields']['观察']}项人工观察文字监督；其中从结构文字截取的观察保留原连续子串与来源，不把关键词索引当异常GT。",
        '- 缺失／冲突字段的值不参与对应token损失，保留JSON结构与EOS监督。空异常框没有被制造为正常负例。完整JSON有效不等于异常描述准确，未认证异常泛化。','',
        '## 下一轮人工复核','']
    for mode,title in [('new','新图类别／关键点／语义'),('existing','已有数据定向语义／联合')]:
        lines.append(f"- **{title}：{counts[mode]}项**，{access[mode]['public_url']} ，沿用原访问码。")
    for mode,label in [('new','新图'),('semantic','已有定向复核')]:
        d=read(ROOT/f'description_inference_{mode}.json')
        lines.append(f"- 全量VLM在{label}队列的实际草稿：完整JSON {d['schema_valid']}/{d['targets']}，其余{d['observation_withheld']}项不采用观察字段，仅保留合法结构恢复及模型原文。这不是自动语义验收通过。")
    lines += [f"- 每细类本轮目标100张不同新图，实际{counts['new']}张，逐类缺口{q['gap']}；不以旧图、模板、重复图或补点凑足。详见`release_quota_report.json`。",
        '- 定向语义队列包含字段冲突、独立分支分歧／格式失败及每类至多10项抽检；已有类别、轮廓和描述作为初始基线保留，新VLM原文只读对照，不自动覆盖人工内容。',
        '- 旧3414／3415入口不替换；新3416／3417使用独立任务、数据库、cookie令牌和本机草稿命名空间。',
        '- 维持右键平移、左键修点、D成功保存后下一条、A不提交、严格K、409冲突保护、原子保存、追加历史、两种独立批量按钮及默认“复核人1”。',
        '- 新入口额外逐次核对来源文件内容哈希，避免秒级文件系统时间戳漏掉同尺寸修改。旧时间戳回归失败及首次命名空间适配失败证据保留；新保护及相关179项CPU回归通过，桌面／390宽实际Chromium操作仅使用临时标注库。',
        '- 空异常描述仍按正常处理，无需逐图填写正常或额外确认。临时Cloudflare网址不保证永久有效，未配置开机自启。','',
        '模型与标签仅为研究标注辅助版本；未覆盖旧训练数据／权重，未晋升正式GT或机器人资格，未自动提交或推送。']
    (ROOT/'REPORT.md').write_text('\n'.join(lines)+'\n')
    write(ROOT/'HANDOFF.json',dict(state='round4_training_and_next_human_review_delivered',sources_current=sources_current(),datasets=data,
        mask_holdout=mask,old_mask_comparison=oldmask,vlm_holdout=vlm,old_vlm_comparison=oldvlm,review_counts=counts,access={k:v['public_url'] for k,v in access.items()},
        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'],
        production_simulation_annotation_POSTs=0,report_sha256=digest(ROOT/'REPORT.md'),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=('audit','report'));a=p.parse_args();(audit_training if a.phase=='audit' else report)()
