"""Round9 research audit: frozen multimodal registry, long-tail/error strata and 320 review candidates.

This script is read-only with respect to existing datasets, SQLite stores and models.
It creates a new research-only directory containing manifests, aggregate statistics and
deterministic candidate pointers; it never writes labels or review events.
"""
from collections import Counter,defaultdict
from datetime import datetime,timezone
import hashlib
import json
from pathlib import Path
import sys

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,jsonl,write,require
from pallet.cluster_targets import digest

DATA=PROJECT/'outputs/pallet_description_round8_training_v1'
OUT=PROJECT/'outputs/pallet_multimodal_research_audit_v3'
TRAINING=DATA/'private/training.jsonl'
MASK=DATA/'mask_run_v1'
VLM=DATA/'vlm_run_v1'


def stable(key,category):return hashlib.sha256(f'{category}:{key}'.encode()).hexdigest()


def load():
    records={r['target_key']:r for r in jsonl(TRAINING)}
    mask_new={r['target_key']:r for r in jsonl(MASK/'independent_branch/holdout_predictions.jsonl')}
    mask_old={r['target_key']:r for r in jsonl(MASK/'old_component_branch_predictions.jsonl')}
    vlm_new={r['target_key']:r for r in jsonl(VLM/'component_branch/holdout_predictions.jsonl')}
    vlm_old={r['target_key']:r for r in jsonl(VLM/'old_component_comparison/holdout_predictions.jsonl')}
    require(set(mask_new)==set(mask_old),'Mask paired holdout mismatch');require(set(vlm_new)==set(vlm_old),'VLM paired holdout mismatch')
    return records,mask_new,mask_old,vlm_new,vlm_old


def evidence(records,mask_new,mask_old,vlm_new,vlm_old):
    candidates=defaultdict(dict)
    def add(category,key,reason,score,evidence):
        if key not in records:return
        old=candidates[category].get(key);value=dict(target_key=key,reason=reason,priority_score=float(score),evidence=evidence)
        if old is None or value['priority_score']>old['priority_score']:candidates[category][key]=value
    for key,p in mask_new.items():
        q=mask_old[key];r=records[key];ref=p['fine_index'];old_ok=q['predicted_fine_index']==ref;new_ok=p['predicted_fine_index']==ref
        if old_ok and not new_ok:add('classification',key,'第八轮旧模型正确而新模型错误',10+abs(q['predicted_fine_index']-p['predicted_fine_index']),dict(reference_index=ref,old_prediction=q['predicted_fine_index'],new_prediction=p['predicted_fine_index'],old_correct=old_ok,new_correct=new_ok))
        elif not old_ok and not new_ok and q['predicted_fine_index']!=p['predicted_fine_index']:add('classification',key,'两模型均错且预测类别改变',5,dict(reference_index=ref,old_prediction=q['predicted_fine_index'],new_prediction=p['predicted_fine_index']))
        if p['iou']<.80:add('geometry',key,'新模型留出IoU低于0.80',1-p['iou'],dict(iou=p['iou'],hd95=p['hd95_native_pixels'],boundary_f1=p['boundary_f1_1px']))
        elif p['iou']<.90:add('geometry',key,'新模型留出IoU位于0.80–0.90风险区间',.90-p['iou'],dict(iou=p['iou'],hd95=p['hd95_native_pixels'],boundary_f1=p['boundary_f1_1px']))
        src=r['source'];flags=[]
        if src.get('foreground_valid_fraction',1)<.75:flags.append('深度有效前景比例低')
        if src.get('range_excursion_to_major_ratio',0)>.35:flags.append('深度范围异常')
        if src.get('geometry_review_flags'):flags.extend(src['geometry_review_flags'])
        if src.get('unit_status') not in ('verified','user_confirmed'):flags.append('单位状态未确认')
        if flags:add('sensor_geometry',key,'；'.join(flags),len(flags),dict(domain=src.get('domain'),unit_status=src.get('unit_status'),scale_to_mm=src.get('scale_to_mm'),foreground_valid_fraction=src.get('foreground_valid_fraction'),geometry_review_flags=src.get('geometry_review_flags',[])))
    for key,p in vlm_new.items():
        q=vlm_old[key];r=records[key];new=p['fields'] if p['valid'] else p.get('recovered_fields');old=q['fields'] if q['valid'] else q.get('recovered_fields');new=new or {};old=old or {}
        changed={f:dict(old=old.get(f),new=new.get(f),reference=r['description'].get({'支腿':'legs','横梁':'beams','材质':'material'}[f])) for f in ('支腿','横梁','材质') if old.get(f)!=new.get(f)}
        if changed:add('vlm',key,'新旧VLM结构字段发生改变',len(changed),dict(changed=changed,old_valid=q['valid'],new_valid=p['valid']))
        src=r['source'];flags=src.get('review_reasons',[])+src.get('geometry_review_flags',[])
        if flags or r['description'].get('observation'):
            add('complex_scene',key,'原人工记录包含场景／几何复核线索',len(flags)+bool(r['description'].get('observation')),dict(review_reasons=flags,observation=r['description'].get('observation'),text=r['description'].get('text')))
    for key,r in records.items():
        src=r['source'];flags=[]
        if src.get('classification_review_required'):flags.append('分类复核标记')
        if src.get('original_appearance_outlier'):flags.append('外观离群')
        if src.get('assigned_class_domain_supported') is False:flags.append('类别域支持不足')
        if src.get('geometry_status')!='ok':flags.append('几何状态非ok')
        if flags:add('annotation_ambiguity',key,'；'.join(flags),len(flags),dict(flags=flags,domain=src.get('domain'),review_reasons=src.get('review_reasons',[]),classification_flags=src.get('classification_review_flags',[])))
    return candidates


def choose(candidates,records,mask_new):
    quotas={'classification':80,'geometry':80,'vlm':60,'sensor_geometry':40,'complex_scene':40,'annotation_ambiguity':20}
    selected=[];used=set();category_counts=defaultdict(Counter)
    for category,budget in quotas.items():
        pool=sorted(candidates[category].values(),key=lambda x:(-x['priority_score'],stable(x['target_key'],category)))
        for item in pool:
            if len([x for x in selected if x['audit_category']==category])>=budget:break
            if item['target_key'] in used:continue
            r=records[item['target_key']];src=r['source'];m=mask_new.get(item['target_key'])
            # Prefer independent components/domains/classes while retaining
            # evidence priority. This avoids turning one capture sequence into
            # the apparent cause of a model error.
            component=item.get('_component',r['component']);domain=src.get('domain','unknown');cls=r['class_id']
            if category_counts[category]['component:'+str(component)]>=2:continue
            if category_counts[category]['domain:'+str(domain)]>=max(8,budget//2):continue
            selected.append(dict(item,audit_category=category,class_id=r['class_id'],parent=r['parent'],split=r['evaluation_branch_split'],component=r['component'],domain=src.get('domain'),
                image=r['image'],target_id=src.get('target_id'),source_pcd_sha256=src.get('source_pcd_sha256'),unit_status=src.get('unit_status'),scale_to_mm=src.get('scale_to_mm'),
                depth_valid_fraction=src.get('foreground_valid_fraction'),mask_iou=m.get('iou') if m else None))
            used.add(item['target_key'])
            category_counts[category]['component:'+str(component)]+=1;category_counts[category]['domain:'+str(domain)]+=1;category_counts[category]['class:'+str(cls)]+=1
        if len([x for x in selected if x['audit_category']==category])<budget:
            # Fill any remaining quota by the highest-priority unused evidence;
            # the diversity pass above is preferred but never drops required n.
            for item in pool:
                if len([x for x in selected if x['audit_category']==category])>=budget:break
                if item['target_key'] in used:continue
                r=records[item['target_key']];src=r['source'];m=mask_new.get(item['target_key'])
                selected.append(dict(item,audit_category=category,class_id=r['class_id'],parent=r['parent'],split=r['evaluation_branch_split'],component=r['component'],domain=src.get('domain'),image=r['image'],target_id=src.get('target_id'),source_pcd_sha256=src.get('source_pcd_sha256'),unit_status=src.get('unit_status'),scale_to_mm=src.get('scale_to_mm'),depth_valid_fraction=src.get('foreground_valid_fraction'),mask_iou=m.get('iou') if m else None));used.add(item['target_key'])
    require(len(selected)==sum(quotas.values()),f'Only selected {len(selected)} of {sum(quotas.values())} audit candidates')
    return selected,quotas


def build():
    require(not OUT.exists(),'Preserve previous research audit; use a new version')
    records,mask_new,mask_old,vlm_new,vlm_old=load();OUT.mkdir(mode=0o700);(OUT/'candidates').mkdir()
    candidates=evidence(records,mask_new,mask_old,vlm_new,vlm_old);selected,quotas=choose(candidates,records,mask_new)
    record_rows=list(records.values());counts=Counter(r['class_id'] for r in record_rows);domains=Counter(r['source'].get('domain','unknown') for r in record_rows);units=Counter((r['source'].get('unit_status'),r['source'].get('scale_to_mm')) for r in record_rows)
    error_summary={cat:dict(available=len(pool),selected=sum(x['audit_category']==cat for x in selected),reasons=dict(Counter(x['reason'] for x in pool.values()))) for cat,pool in candidates.items()}
    write(OUT/'dataset_registry.json',dict(state='round9_readonly_multimodal_registry_complete',targets=len(records),distinct_images=len({r['image_sha256'] for r in record_rows}),fine_classes=len(counts),parents=len({r['parent'] for r in record_rows}),splits=dict(Counter(r['evaluation_branch_split'] for r in record_rows)),semantic_supervision=8425,observation_supervision=1039,class_counts=dict(counts),domains=dict(domains),unit_status=dict((f'{k[0]}|{k[1]}',v) for k,v in units.items()),source_manifest_sha256=digest(DATA/'dataset_manifest.json'),source_binding_sha256=digest(DATA/'source_binding.json'),no_existing_files_or_databases_modified=True))
    write(OUT/'error_strata.json',dict(state='round9_deterministic_error_strata_complete',categories=error_summary,model_pairing=dict(mask_holdout=len(mask_new),vlm_holdout=len(vlm_new)),definitions={
        'classification':'paired old/new class output errors', 'geometry':'paired mask IoU and boundary risk', 'vlm':'paired structure-field or validity changes',
        'sensor_geometry':'depth validity/unit/geometry evidence flags', 'complex_scene':'human observation/review reasons', 'annotation_ambiguity':'source classification/geometry/outlier flags'}))
    write(OUT/'audit_candidates.jsonl',selected,lines=True)
    for cat in quotas:
        write(OUT/'candidates'/f'{cat}.jsonl',[x for x in selected if x['audit_category']==cat],lines=True)
    write(OUT/'audit_plan.json',dict(state='round9_320_candidate_plan_not_human_labels',total=len(selected),quotas=quotas,selection='priority_then_stable_hash_no_duplicate_target',human_review_required=True,annotation_POSTs=0))
    write(OUT/'literature_positioning.json',dict(state='research_positioning_compiled',references=[
        dict(name='RGB-D pallet detection and fork insertion',year=2021,doi='10.1109/ICMA52036.2021.9512641',relevance='RGB-D语义分割+深度位姿+插取验证，说明单纯ToF/几何融合不足以构成独立创新'),
        dict(name='Pallet Detection and Estimation with RGB-D Salient Feature Learning',year=2023,doi='10.1109/CAC59555.2023.10451178',relevance='低成本ToF、RGB/深度显著特征和点云匹配，作为强基线'),
        dict(name='Candle: Efficient and Long-Tailed Generalization for Pre-trained VLM',year=2024,arxiv='2406.12638',relevance='长尾、base/new类和跨模态原型，适合作为Prompt/长尾对照'),
        dict(name='Exploring Vision-Language Models for Imbalanced Learning',year=2023,arxiv='2304.01457',relevance='Prompt tuning、Balanced Softmax/Focal和尾类指标'),
        dict(name='Generalized OOD Detection in the VLM Era',year=2024,arxiv='2407.21794',relevance='区分near-OOD、开放集和工业异常，避免把未知构型等同普通分类错误')],claim_boundary='不宣称首次RGB-D/ToF托盘检测；创新应落在证据约束多模态、长尾Prompt和叉取可行性开放集决策'))
    report=['# 第九轮研究准备：多模态、长尾与异常审计','',
        f"冻结口径：{len(records)}目标／{len({r['image_sha256'] for r in record_rows})}原图／{len(counts)}细类／{len({r['parent'] for r in record_rows})}大类；训练{sum(r['evaluation_branch_split']=='train' for r in record_rows)}、组件留出{sum(r['evaluation_branch_split']=='holdout' for r in record_rows)}。",'',
        '本阶段只读现有数据和模型输出，不写SQLite、不生成correct事件、不改正式标签。',
        f"自动错误分层共覆盖{sum(len(v) for v in candidates.values())}个去重候选池，实际确定320项：",
        '；'.join(f'{k}{v}项' for k,v in quotas.items())+'。每个目标只进入一个审计类别，后续需要人工复核。','',
        '## 已识别的论文问题','',
        '- ToF多域原始尺度并不统一：未标定域不应笼统宣称毫米级精度；先做域级标定和有效深度统计。',
        '- 第八轮细类和大类准确率回退、IoU略升；VLM材质改善但支腿／横梁回退，说明需要逐错误来源分析而不是继续堆叠模型。',
        '- 长尾类别、未知构型、深度无效、遮挡、出视野和非托盘场景应分别作为错误类型和拒识对象。','',
        '## 下一阶段实验顺序','',
        '1. 320项人工错误审计：分类80、几何80、VLM60、传感器/几何40、复杂场景40、标注歧义20；至少25%双人复核。',
        '2. 多模态基线：强度、深度、有效性、早期融合、双编码器门控融合；输出类别、轮廓、结构锚点、置信度和Unknown。',
        '3. 长尾Prompt：Head/Medium/Tail分层，1/5/10-shot，属性组合Prompt、Prompt ensemble、Balanced Softmax和尾类原型对照。',
        '4. OOD/风险覆盖：未知构型、非托盘、深度异常和复杂遮挡，报告AUROC、FPR95、OSCR、Risk-Coverage和不安全场景漏拒率。',
        '5. 叉取决策：VLM只提供场景证据，几何约束负责硬门控，最终输出可叉取／不可叉取／无法判断；实机成功率另行验证。','',
        '详细候选、数据注册、文献定位和来源哈希见本目录JSONL/JSON文件。下一步执行应先完成320项人工审计，再决定关键点和OOD标注协议。']
    (OUT/'REPORT.md').write_text('\n'.join(report)+'\n')
    write(OUT/'HANDOFF.json',dict(state='round9_research_audit_phase1_complete',dataset_registry_sha256=digest(OUT/'dataset_registry.json'),error_strata_sha256=digest(OUT/'error_strata.json'),candidate_plan_sha256=digest(OUT/'audit_plan.json'),candidates=320,production_annotation_POSTs=0,existing_databases_modified=False,next_gate='human_320_error_audit_before_multimodal_training'))
    print(dict(state='round9_phase1_complete',output=str(OUT.relative_to(PROJECT)),candidates=len(selected),quotas=quotas))


if __name__=='__main__':build()
