"""Replay latest annotations and actual training, then issue a bounded delivery report."""
from contextlib import closing
from datetime import datetime,timezone
import json
from pathlib import Path
import sys

import cv2
import numpy as np

PROJECT=Path(__file__).resolve().parents[1];sys.path.insert(0,str(PROJECT));sys.path.insert(0,str(PROJECT/'scripts'))
from pallet.mask_model_training_data import read,write,jsonl,require
from pallet.cluster_targets import digest
from pallet.mask_template_refined import readonly
from pallet.pose_review_geometry import raster_mask
from pallet.pallet_description_round5_data import ROOT,PREVIOUS,SOURCE_SETS,verify,sources_current
from pallet.pallet_description_round5_release import REVIEW_SETS,RELEASE,description_draft
from pallet.pallet_description_round5_vlm import format_target,parse


def source_replay():
    _,records=verify();baseline={r['target_key']:r for r in jsonl(PREVIOUS/'private/training.jsonl')};oldkeys=set(baseline);scopes={}
    for mode,s in SOURCE_SETS.items():
        props={p['target_key']:p for p in jsonl(s['root']/'proposals.jsonl')}
        with closing(readonly(ROOT/f'private/snapshots/{mode}.sqlite3')) as db:
            require(db.execute('PRAGMA quick_check').fetchall()==[('ok',)],'Snapshot damaged')
            current={k:json.loads(raw) for k,raw in db.execute('SELECT target_key,payload FROM reviews')};last={}
            for k,rev,raw in db.execute('SELECT target_key,revision,payload FROM events ORDER BY id'):
                require(rev==last.get(k,{}).get('revision',0)+1,'Noncontiguous history');last[k]=json.loads(raw)
        require(last==current and set(props)==set(current) and all(r['action']=='correct' for r in current.values()),'Completed source history mismatch')
        for key,p in props.items():
            v=current[key];baseline[key]=dict(**{f:v[f] for f in ('polygon','class_id','description')},image_sha256=p['image_sha256'])
        scopes[mode]=dict(targets=len(current),events=sum(v['revision'] for v in current.values()))
    require(set(baseline)=={r['target_key'] for r in records},'Target merge coverage mismatch')
    for r in records:
        ref=baseline[r['target_key']];require(all(r[f]==ref[f] for f in ('polygon','class_id','description','image_sha256')),'Latest human record overwritten or omitted')
        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(),'Mask clipped or target supervision invalid')
    require(sources_current(),'Sources changed during independent replay')
    write(ROOT/'source_replay.json',dict(state='round5_completed_latest_labels_and_native_masks_replayed',targets=len(records),new_targets=len(set(baseline)-oldkeys),sources=scopes,
        current_equals_history=True,source_actions_not_rewritten=True,production_annotation_POSTs=0))


def training():
    import torch
    require(read(ROOT/'source_replay.json')['state']=='round5_completed_latest_labels_and_native_masks_replayed','Source replay required')
    _,records=verify();mask=ROOT/'mask_run_v1';vlm=ROOT/'vlm_run_v1';complete=read(mask/'training_complete.json');vcomplete=read(vlm/'training_complete.json');steps={}
    for phase in ('independent_branch','full_pool'):
        folder=mask/phase;receipt=read(folder/'training_receipt.json');weights=torch.load(folder/'final_ema.pt',map_location='cpu',weights_only=True)
        expected={r['target_key'] for r in records if phase=='full_pool' or r['evaluation_branch_split']=='train'}
        require(set(weights['trained_targets'])==expected and weights['model']['fine_head.weight'].shape[0]==40,'Wrong actual trained targets/head dimensions')
        history=[read(p) for p in sorted(folder.glob('epoch_*.json'))]
        require(receipt['optimization_steps']==weights['optimization_steps']==sum(v['epoch_updates'] for v in history)>0,'No real mask optimizer history')
        require(receipt['checkpoint_sha256']==digest(folder/'final_ema.pt')==complete['full_pool' if phase=='full_pool' else 'independent_branch'],'Mask hash changed')
        require(all(np.isfinite(list(v['mean_losses'].values())).all() for v in history),'Nonfinite mask training');steps['mask_'+phase]=receipt['optimization_steps']
    sem=jsonl(ROOT/'description_data/records.jsonl');eligible=[r for r in sem if any(r['supervised_fields'].values())]
    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')
        expected=[r for r in eligible if phase=='full_pool' or r['split']=='train']
        require(receipt['records']==len(expected) and receipt['optimizer_steps']>0 and receipt['adapter_sha256']==digest(folder/'adapter/adapter_model.safetensors'),'Actual VLM optimizer/data binding differs')
        require(plan['independent_base_start']==(phase=='component_branch'),'Independent VLM branch contaminated by full-pool checkpoint')
        require(all(r['eos_supervised'] and (not r['missing_observation_has_zero_semantic_characters'] or r['observation_quote_tokens_supervised']) for r in audit),'Missing observation quotes/EOS masked')
        steps['vlm_'+phase]=receipt['optimizer_steps']
    require(vcomplete['adapter_sha256']==digest(vlm/'full_pool/adapter/adapter_model.safetensors'),'Full VLM changed')
    require(all(not r['missing_observation_is_normal'] and (r['supervised_fields']['观察'] or format_target(r)['观察']=='') for r in sem),'Missing observation became normal GT')
    require(sources_current(),'Source changed during training audit')
    write(ROOT/'training_verification.json',dict(state='round5_actual_training_and_latest_source_replay_passed',optimizer_steps=steps,targets=len(records),
        VLM_targets=len(eligible),dataset_manifest_sha256=digest(ROOT/'dataset_manifest.json'),no_fabricated_correct_events=True,no_normal_negative_labels_inferred=True,
        actual_missing_observation_quotes_and_EOS_supervised=True,verified_at=datetime.now(timezone.utc).isoformat()))


def bundles():
    baseline={r['target_key']:r for r in jsonl(ROOT/'private/training.jsonl')};protected=set(read(ROOT/'private/protection.json')['protected_image_hashes']);results={}
    for mode,s in REVIEW_SETS.items():
        props=jsonl(s['root']/'proposals.jsonl');m=read(s['root']/'manifest.json');preds={p['target_key']:p for p in jsonl(ROOT/f'descriptions_{"new" if mode=="new" else "semantic"}.jsonl')}
        require(len(props)==len({p['target_key'] for p in props})==m['request_count'] and m['proposals_sha256']==digest(s['root']/'proposals.jsonl'),'Wrong task identities')
        for p in props:
            pred=preds[p['target_key']];require(p['description_raw']==pred['raw'] and parse(pred['raw'])['valid']==pred['valid'],'Raw model output not replayed')
            if mode=='new':require(p['target_key'] not in baseline and p['image_sha256'] not in protected and p['description']==description_draft(pred),'Not a new original or model draft altered')
            else:require(all(p[f]==baseline[p['target_key']][f] for f in ('polygon','class_id','description')),'Existing human baseline overwritten')
        if mode=='new':require(len(props)==len({p['image_sha256'] for p in props}),'Repeated new image')
        for n,h in m['public_assets'].items():require(digest(s['root']/n)==h,'Task asset changed')
        results[mode]=dict(targets=len(props),unique_images=len({p['image_sha256'] for p in props}),manifest_sha256=digest(s['root']/'manifest.json'))
    require(sources_current(),'Sources changed during bundle replay');write(ROOT/'bundle_source_replay.json',dict(state='round5_next_tasks_latest_human_baseline_and_model_raw_replayed',bundles=results,production_annotation_POSTs=0))


def report():
    require(read(ROOT/'training_verification.json')['state']=='round5_actual_training_and_latest_source_replay_passed','Actual training verification required')
    require(read(ROOT/'bundle_source_replay.json')['state']=='round5_next_tasks_latest_human_baseline_and_model_raw_replayed','Task 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')
    vrecovery=read(ROOT/'vlm_run_v1/component_branch/structure_recovery_audit.json');orecovery=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')};vp=jsonl(ROOT/'vlm_run_v1/component_branch/holdout_predictions.jsonl')
    positive=[p for p in vp if semantic[p['target_key']]['observation_supervised']]
    observation_answered=sum(p['valid'] and bool(p['fields'].get('观察')) for p in positive)
    counts={};urls={}
    for mode,s in REVIEW_SETS.items():
        require(read(RELEASE/mode/'round5_browser_verification.json')['state']=='passed_round5_actual_task_desktop_and390_atomic_batch_and_fast_navigation','Actual UI proof required')
        p=read(s['managed']/'round5_public_verification.json');require(p['state']=='round5_authenticated_public_and_fast_guard_passed','Public readback required')
        counts[mode]=read(s['root']/'manifest.json')['request_count'];urls[mode]=p['public_url']
    lines=['# 第五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']}。",'',
        '## 同一份最新标签组件留出对比','', '| 指标 | 上轮独立模型 | 本轮独立模型 |','|---|---:|---:|',
        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} |",'',
        f"训练侧缺少细类{mask['classes_absent_from_training']}，这些细类不能声称已由独立分支验证。完整标注模型随后使用全部40类。沿用同图／采集关联分量划分，旧独立分支训练目标与本次留出无交集，但样本曾由旧全量模型筛选，且反复用于迭代，不能当新采集盲测。",'',
        '## VLM格式修正与语义边界','',f"完整JSON：旧{oldvlm['strict_complete_JSON_valid']}/{oldvlm['targets']} → 新{vlm['strict_complete_JSON_valid']}/{vlm['targets']}。各自冻结提示词，同样256 token预算；不是仅改变单一因素的消融。",
        '- 缺失观察以前把整个null值的token掩掉，模型没有学到该位置合法值的格式。本轮统一观察为带双引号字符串；缺失项没有语义字符，仅监督引号／JSON结构／EOS。空字符串表示缺少可靠观察信息，不是正常GT。',
        f"- 观察文字监督{data['semantic_supervised_fields']['观察']}项，保留异常框原文或结构文字连续原子串及来源；关键词索引不是异常GT。",
        '- 完整JSON有效仍不等于描述准确；以下结构一致率含合法结构恢复，分母为全部有对应标签的留出项：']
    for k in ('支腿','横梁','材质'):lines.append(f"  - {k}：旧{orecovery['attribute_exact_including_explicit_structure_recovery'][k]['accuracy']:.2%} → 新{vrecovery['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，{observation_answered}项输出非空观察，{len(positive)-observation_answered}项为空／不可用；非空不代表内容正确。模型确有较多弃答，不能把完整JSON通过当作可靠异常检出率。",
        '- 异常泛化尚未认证，没有从空异常框制造正常负例。全量模型草稿需人工复核。界面已明确“模型空观察为未知”，不改变用户要求的应用空异常默认正常规则。','', '## 下一轮入口','']
    for mode,title in [('new','新图类别／轮廓／语义'),('existing','已有数据定向语义与联合复核')]:lines.append(f"- **{title}：{counts[mode]}项**，{urls[mode]} ，沿用原访问码。")
    for mode in ('new','semantic'):
        d=read(ROOT/f'description_inference_{mode}.json');lines.append(f"- {mode}实际全量VLM输出：完整JSON {d['schema_valid']}/{d['targets']}，格式无效观察拒绝{d['observation_withheld']}；有效但观察为空／未知{d['strict_valid_empty_observation']}。")
    selection=read(ROOT/'release_selection.json')
    lines += [f"- 新图每类100张目标4000，实际{counts['new']}，逐类缺口{selection['gap']}；未混入旧图、复制模板、补点或放宽质量门槛。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，新模型原文只读对照；包含独立分支分歧／格式失败和每类至多10项抽检，不当作新的配额图。',
        '- 新3418／3419使用独立任务和草稿命名空间；已完成3416／3417及所有旧库、旧模型保持。',
        '- 默认复核人1，空异常无需填写正常；保留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、原子追加历史与两种独立批量按钮。沿用并行全内容哈希加速，不缓存人工revision、不用TTL或内容抽样。211项CPU回归通过，含实际token引号监督、本机旧草稿、原子批量和来源失效保护。',
        '- 桌面／390宽真实Chromium验收仅在临时库进行，公网鉴权／资源／逐类详情只读核验；未做真实MacOS/iPhone Safari实机验收。网址为临时Cloudflare，未配置开机自启。','',
        '本轮只新增版本化研究训练、模型与人工复核任务；未覆盖正式标签、旧训练数据／权重／历史，未晋升GT或机器人资格，无生产模拟标注POST，未提交或推送。']
    require(sources_current(),'Sources changed before final report');(ROOT/'REPORT.md').write_text('\n'.join(lines)+'\n')
    write(ROOT/'HANDOFF.json',dict(state='round5_real_training_and_next_review_delivered',datasets=data,mask_holdout=mask,old_mask=oldmask,vlm_holdout=vlm,old_vlm=oldvlm,
        review_counts=counts,access=urls,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'));a=p.parse_args()
    {'source':source_replay,'training':training,'bundles':bundles,'report':report}[a.phase]()
