"""Assign appearance outliers to frozen centers of the user's 22 merged classes."""
from __future__ import annotations

import argparse
from collections import Counter
import html
import io
import json
from pathlib import Path
import re

import numpy as np

from pallet.fork_face_anylabeling import digest, encoded, protected_write, safe_path
from pallet.merge_appearance_classes import read_json, read_jsonl, relative, write_csv, write_workbook


DEFAULT_MERGED = Path('outputs/pallet_appearance_class_merge_v1')
DEFAULT_CLUSTERING = Path('outputs/pallet_target_clustering_v1/analysis/run_006')
DEFAULT_OUTPUT = Path('outputs/pallet_appearance_center_reclassification_v1')


def classify_features(features, labels, radius_quantile=.95, margin_threshold=.10):
    """Sample-weighted arithmetic centers; outliers never participate in fitting."""
    x=np.asarray(features,dtype=np.float64);labels=np.asarray(labels)
    if (x.ndim!=2 or labels.shape!=(len(x),) or not np.isfinite(x).all()
            or labels.dtype.kind not in 'iu' or (labels==0).any() or (labels < -1).any()):
        raise ValueError('Invalid finite appearance features or merged class labels')
    if not 0<radius_quantile<1 or not 0<=margin_threshold<1:
        raise ValueError('Invalid reference radius quantile or distance margin threshold')
    ids=np.array(sorted(set(labels)-{-1}),dtype=np.int64)
    if len(ids)<2:raise ValueError('At least two nonempty reference classes are required')
    centers=np.stack([x[labels==c].mean(axis=0) for c in ids])
    counts=np.array([int((labels==c).sum()) for c in ids])
    distance=np.linalg.norm(x[:,None,:]-centers[None,:,:],axis=2)
    order=np.argsort(distance,axis=1,kind='stable')
    nearest=order[:,0];second=order[:,1];index=np.arange(len(x))
    d1=distance[index,nearest];d2=distance[index,second]
    # Equal zero distances are a tie and receive margin zero, not certainty.
    margin=np.divide(d2-d1,d2,out=np.zeros_like(d1),where=d2>1e-12)
    radii=np.array([np.quantile(distance[labels==c,j],radius_quantile) for j,c in enumerate(ids)])
    final_labels=np.where(labels==-1,ids[nearest],labels)
    class_positions={int(c):j for j,c in enumerate(ids)}
    final_positions=np.array([class_positions[int(c)] for c in final_labels])
    assigned_distance=distance[index,final_positions]
    far=assigned_distance>radii[final_positions]
    ambiguous=margin<margin_threshold
    return {'class_ids':ids,'centers':centers,'reference_counts':counts,'radii':radii,
        'distances':distance,'nearest_positions':nearest,'second_positions':second,
        'nearest_class':ids[nearest],'second_class':ids[second], 'nearest_distance':d1,'second_distance':d2,
        'margin':margin,'final_labels':final_labels,'final_positions':final_positions,
        'assigned_distance':assigned_distance,'far':far,'ambiguous':ambiguous,
        'priority_review':(labels==-1)&(far|ambiguous), 'reassigned':labels==-1}


def align_features(rows, source_rows, indices, features):
    """NPZ indices refer to run_006 row positions, never to an assumed ID ordering."""
    indices=np.asarray(indices);features=np.asarray(features)
    if (indices.ndim!=1 or indices.dtype.kind not in 'iu' or len(indices)!=len(features)
            or len(set(indices.tolist()))!=len(indices) or (indices<0).any() or (indices>=len(source_rows)).any()):
        raise ValueError('Invalid or duplicate cached appearance indices')
    positions={source_rows[int(index)]['target_id']:j for j,index in enumerate(indices)}
    if len(positions)!=len(indices):raise ValueError('Duplicate cached source target identity')
    source_by_id={r['target_id']:r for r in source_rows}
    if len(source_by_id)!=len(source_rows) or {r['target_id'] for r in rows}!=set(source_by_id):
        raise ValueError('Merged target identities disagree with source clustering')
    for row in rows:
        raw=source_by_id[row['target_id']]
        if any(row.get(k)!=v for k,v in raw.items()):raise ValueError('Original run_006 target metadata changed')
    missing=[r['target_id'] for r in rows if r['target_id'] not in positions]
    if missing:raise ValueError(f'Appearance feature missing for {len(missing)} targets')
    return np.asarray(features[[positions[r['target_id']] for r in rows]],dtype=np.float64)


def enrich_rows(rows, result, group_by_id):
    output=[]
    for i,row in enumerate(rows):
        new=bool(result['reassigned'][i]);class_id=int(result['final_labels'][i]);position=result['final_positions'][i]
        radius=float(result['radii'][position]);distance=float(result['assigned_distance'][i])
        flags=[]
        if new and result['far'][i]:flags.append('beyond_reference_p95')
        if new and result['ambiguous'][i]:flags.append('small_top2_margin')
        value=dict(row)
        value.update(final_class_id=class_id,final_class=group_by_id[class_id]['merged_class'],
            classification_source='nearest_merged_centroid' if new else 'retained_excel_merge',
            original_appearance_outlier=new,
            nearest_center_class_id=int(result['nearest_class'][i]),
            second_center_class_id=int(result['second_class'][i]),
            nearest_center_distance=float(result['nearest_distance'][i]),
            second_center_distance=float(result['second_distance'][i]),
            assigned_center_distance=distance, reference_distance_p95=radius,
            distance_to_reference_p95_ratio=distance/radius if radius>0 else (0. if distance==0 else None),
            centroid_top2_margin=float(result['margin'][i]),
            classification_review_required=new,classification_priority_review=bool(result['priority_review'][i]),
            classification_review_flags=flags,
            assigned_class_domain_supported=row['domain'] in group_by_id[class_id]['domains'],
            reference_center_disagreement=not new and int(result['nearest_class'][i])!=class_id)
        output.append(value)
    return output


def aggregate(groups, rows):
    output=[]
    for group in groups:
        selected=[r for r in rows if r['final_class_id']==group['merged_class_id']]
        added=[r for r in selected if r['original_appearance_outlier']]
        output.append({**group,'reference_targets':group['targets'],'added_outlier_targets':len(added),
            'total_targets':len(selected),'priority_review_targets':sum(r['classification_priority_review'] for r in added),
            'far_from_center_targets':sum('beyond_reference_p95' in r['classification_review_flags'] for r in added),
            'ambiguous_targets':sum('small_top2_margin' in r['classification_review_flags'] for r in added),
            'new_domain_targets':sum(not r['assigned_class_domain_supported'] for r in added),
            'final_splits':{s:sum(r['split']==s for r in selected) for s in ('train','val','test')},
            'added_splits':{s:sum(r['split']==s for r in added) for s in ('train','val','test')},
            'final_frames':len({r['frame_id'] for r in selected}),
            'final_acquisition_groups':len({r['group_id'] for r in selected})})
    return output


STYLE='''@font-face{font-family:CenterReport;src:url(assets/ReportSans.woff)}*{box-sizing:border-box}body{margin:0;background:#f0f4f7;color:#22384d;font:15px/1.75 CenterReport,"Microsoft YaHei",sans-serif}main{max-width:1360px;margin:auto;padding:32px 24px}h1{font-size:30px}h2{margin-top:32px}a{color:#006e86}table{border-collapse:collapse;width:100%;background:white}th,td{padding:10px 12px;border-bottom:1px solid #dce4ec;text-align:left;font-variant-numeric:tabular-nums}th{background:#173e59;color:white;white-space:nowrap}tr:nth-child(even){background:#f8fafb}.table{overflow:auto;border:1px solid #dbe4eb;border-radius:9px}.banner{padding:20px 24px;border-radius:12px;background:#173e59;color:white}.filters,.links{display:flex;gap:12px;flex-wrap:wrap;margin:20px 0}select,input,button{font:inherit;padding:8px 11px;border:1px solid #c6d3df;border-radius:7px;background:white}.cards,.examples{display:grid;grid-template-columns:repeat(auto-fill,minmax(min(270px,100%),1fr));gap:16px}.card,.example{background:white;border:1px solid #dbe4eb;border-radius:10px;padding:12px;cursor:pointer}.card img,.example img{height:120px;width:100%;object-fit:contain;background:#182936;image-rendering:pixelated}.card small,.example small{display:block;overflow-wrap:anywhere}.tag{font-size:13px;color:#58728a}.priority{color:#ad5b13}.pages{display:flex;gap:16px;justify-content:center;align-items:center;margin:24px 0}.group{margin:24px 0;padding:18px;background:#e8eff4;border-radius:12px}dialog{border:0;border-radius:12px;width:min(1050px,95vw);max-height:94vh}dialog::backdrop{background:#102b3dbb}dialog canvas{max-width:100%;height:auto;image-rendering:pixelated}code,pre{font-family:CenterReport,ui-monospace,monospace;overflow-wrap:anywhere}pre{white-space:pre-wrap}.comparison{display:flex;gap:15px;flex-wrap:wrap}.comparison img{width:250px;height:105px;object-fit:contain;background:#182936}@media(max-width:600px){main{padding:20px 14px}h1{font-size:25px}}'''


def gallery(ctx, rows, groups, center_records):
    output,clustering=ctx['output'],ctx['clustering']
    fields=['target_id','target_key','frame_id','box_index','group_id','domain','split','image','bbox',
        'original_appearance_class','appearance_cluster','final_class_id','final_class','classification_source',
        'assigned_center_distance','nearest_center_class_id','second_center_class_id','nearest_center_distance',
        'second_center_distance','distance_to_reference_p95_ratio','centroid_top2_margin','classification_review_required',
        'classification_priority_review','classification_review_flags','assigned_class_domain_supported']
    data=[{k:r[k] for k in fields} for r in rows]
    template='''<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>离群目标按新中心重新分类</title><style>__STYLE__</style><main><h1>22 类中心 · 离群目标重新分类</h1><p class="links"><a href="report_zh.html">结果报告与对比样本</a><a href="离群目标重新分类结果.xlsx">Excel统计</a><a href="outlier_assignments.csv">离群目标清单</a><a href="targets.csv">全量分类清单</a></p><div class="banner">5,362 个原离群目标已分配至 22 类 · 6,982 个原归属保留 · 新归类均待人工复核</div><div class="filters"><select id="class" aria-label="最终类别"><option value="">全部22类</option></select><select id="source" aria-label="归类来源"><option value="new" selected>本次离群重新分类</option><option value="all">全部目标</option><option value="reference">保留的原类别</option></select><select id="review" aria-label="复核依据"><option value="">全部复核状态</option><option value="priority">优先复核</option><option value="standard">常规复核</option><option value="far">超过参考P95距离</option><option value="ambiguous">前两候选接近</option><option value="newdomain">参考类未覆盖的配置域</option></select><select id="split" aria-label="原划分"><option value="">全部原划分</option><option>train</option><option>val</option><option>test</option></select><select id="domain" aria-label="配置域"><option value="">全部配置域</option></select><select id="sort" aria-label="排序"><option value="original">原始顺序</option><option value="far">中心距离从大到小</option><option value="ratio">相对P95距离从大到小</option><option value="ambiguous">前两类间隔从小到大</option></select><input id="search" placeholder="搜索目标或采集组" aria-label="搜索"><span id="count"></span></div><p>距离和间隔是特征空间诊断值，不是分类正确率或概率。</p><div id="cards" class="cards"></div><div class="pages"><button id="prev">上一页</button><span id="page"></span><button id="next">下一页</button></div></main><dialog id="detail"><button id="close">关闭</button><h2 id="title"></h2><canvas id="image"></canvas><div class="comparison"><div><p>最近类别的原参考样本</p><img id="first-example" alt="最近中心参考样本"></div><div><p>第二近类别的原参考样本</p><img id="second-example" alt="第二近中心参考样本"></div></div><p class="links"><a id="original" target="_blank">原生强度图</a><a id="original-gallery" target="_blank">原聚类详情</a></p><pre id="metadata"></pre></dialog><script>
const rows=__DATA__,groups=__GROUPS__,centers=__CENTERS__,media=__MEDIA__,dataset=__DATASET__,oldGallery=__OLD__, $=id=>document.getElementById(id);let page=0,current=null;const pageSize=48;
function option(id,value,label){const o=document.createElement('option');o.value=value;o.textContent=label;$(id).append(o)}for(const g of groups)option('class',String(g.merged_class_id),`${g.merged_class} · 新增${g.added_outlier_targets} / 总计${g.total_targets}`);for(const d of [...new Set(rows.map(r=>r.domain))].sort())option('domain',d,d);
function isNew(r){return r.classification_source==='nearest_merged_centroid'}function flags(r){return r.classification_review_flags}function render(){const term=$('search').value.trim().toLowerCase();let selected=rows.filter(r=>(!$('class').value||String(r.final_class_id)===$('class').value)&&($('source').value==='all'||($('source').value==='new'?isNew(r):!isNew(r)))&&(!$('split').value||r.split===$('split').value)&&(!$('domain').value||r.domain===$('domain').value)&&(!term||(r.target_id+' '+r.group_id).toLowerCase().includes(term))&&({priority:r.classification_priority_review,standard:isNew(r)&&!r.classification_priority_review,far:flags(r).includes('beyond_reference_p95'),ambiguous:flags(r).includes('small_top2_margin'),newdomain:isNew(r)&&!r.assigned_class_domain_supported}[$('review').value]??true));if($('sort').value==='far')selected.sort((a,b)=>b.assigned_center_distance-a.assigned_center_distance);if($('sort').value==='ratio')selected.sort((a,b)=>(b.distance_to_reference_p95_ratio??Infinity)-(a.distance_to_reference_p95_ratio??Infinity));if($('sort').value==='ambiguous')selected.sort((a,b)=>a.centroid_top2_margin-b.centroid_top2_margin);const pages=Math.max(1,Math.ceil(selected.length/pageSize));page=Math.min(page,pages-1);$('cards').replaceChildren();$('count').textContent=`${selected.length.toLocaleString()} 个目标`;for(const r of selected.slice(page*pageSize,(page+1)*pageSize)){const card=document.createElement('article');card.className='card';const im=document.createElement('img');im.src=media+'/'+r.target_key+'.jpg';im.alt=r.target_id;im.loading='lazy';const title=document.createElement('b');title.textContent=`${r.final_class} · ${isNew(r)?'原离群':r.original_appearance_class}`;const quality=document.createElement('small');quality.className=r.classification_priority_review?'priority':'tag';quality.textContent=`距离 ${r.assigned_center_distance.toFixed(3)} · 间隔 ${r.centroid_top2_margin.toFixed(3)} · ${r.classification_priority_review?'优先复核':(isNew(r)?'待复核':'原归属保留')}`;const meta=document.createElement('small');meta.textContent=r.split+' · '+r.domain;const id=document.createElement('small');id.textContent=r.frame_id+' · 框'+r.box_index;card.append(im,title,quality,meta,id);card.onclick=()=>openTarget(r);$('cards').append(card)}$('page').textContent=`${page+1} / ${pages}`;$('prev').disabled=page===0;$('next').disabled=page+1===pages}
function openTarget(r){current=r.target_id;$('detail').showModal();$('title').textContent=r.final_class+' · '+(isNew(r)?'本次离群重新归类':r.original_appearance_class);const image=new Image();image.onload=()=>{if(current!==r.target_id)return;const c=$('image');c.width=image.width;c.height=image.height;const cx=c.getContext('2d');cx.drawImage(image,0,0);cx.strokeStyle='#ff4c45';cx.lineWidth=Math.max(1,image.width/250);cx.strokeRect(r.bbox.x,r.bbox.y,r.bbox.width,r.bbox.height)};image.src=dataset+'/'+r.image;$('original').href=image.src;$('original-gallery').href=oldGallery+'?target='+encodeURIComponent(r.target_key);$('first-example').src=media+'/'+centers.find(c=>c.class_id===r.nearest_center_class_id).representative_target_key+'.jpg';$('second-example').src=media+'/'+centers.find(c=>c.class_id===r.second_center_class_id).representative_target_key+'.jpg';$('metadata').textContent=JSON.stringify({最终类别:r.final_class,原A类:r.original_appearance_class,目标:r.target_id,原划分:r.split,采集组:r.group_id,配置域:r.domain,最近类别:r.nearest_center_class_id,最近距离:r.nearest_center_distance,第二类别:r.second_center_class_id,第二距离:r.second_center_distance,前两类相对间隔:r.centroid_top2_margin,相对参考P95距离:r.distance_to_reference_p95_ratio,优先复核原因:flags(r),参考类覆盖该配置域:r.assigned_class_domain_supported,原框:r.bbox},null,2)}
for(const id of ['class','source','review','split','domain','sort','search'])$(id).addEventListener(id==='search'?'input':'change',()=>{page=0;render()});$('prev').onclick=()=>{page--;render()};$('next').onclick=()=>{page++;render()};$('close').onclick=()=>$('detail').close();const query=new URLSearchParams(location.search);if(query.has('class'))$('class').value=query.get('class');if(query.has('source'))$('source').value=query.get('source');if(query.has('review'))$('review').value=query.get('review');render();if(query.has('target')){const r=rows.find(r=>r.target_key===query.get('target'));if(r)openTarget(r)}
</script></html>'''
    replacements={'__STYLE__':STYLE,'__DATA__':json.dumps(data,ensure_ascii=False,separators=(',',':')).replace('<','\\u003c'),
        '__GROUPS__':json.dumps(groups,ensure_ascii=False,separators=(',',':')).replace('<','\\u003c'),
        '__CENTERS__':json.dumps(center_records,ensure_ascii=False,separators=(',',':')).replace('<','\\u003c'),
        '__MEDIA__':json.dumps(relative(clustering.parent.parent/'media',output)),
        '__DATASET__':json.dumps(relative(ctx['dataset'],output)),
        '__OLD__':json.dumps(relative(clustering/'gallery.html',output))}
    protected_write(output/'gallery.html',re.sub('|'.join(map(re.escape,replacements)),lambda m:replacements[m[0]],template).encode())


def reports(ctx, rows, groups, centers, summary):
    from markdown_it import MarkdownIt
    table=['| 新类别 | 原参考目标 | 新增离群目标 | 最终总数 | 优先复核 | train | val | test |',
           '|---|---:|---:|---:|---:|---:|---:|---:|']
    for g in groups:table.append(f'| {g["merged_class"]} | {g["reference_targets"]} | {g["added_outlier_targets"]} | {g["total_targets"]} | {g["priority_review_targets"]} | {g["final_splits"]["train"]} | {g["final_splits"]["val"]} | {g["final_splits"]["test"]} |')
    lines=['# 基于22个新类别中心的离群目标重新分类','',
        f'将剩余 **{summary["reassigned_outliers"]:,} 个原外观离群目标**分配到最近的新类别中心，原 **{summary["retained_reference_targets"]:,} 个目标归属保留**。最终 {summary["total_targets"]:,} 个目标均有 Class 1–Class 22 分组。','',
        '[打开可筛选图册](gallery.html) · [下载Excel统计](离群目标重新分类结果.xlsx) · [离群分类CSV](outlier_assignments.csv) · [完整目标CSV](targets.csv) · [优先复核CSV](priority_review.csv)','',
        '## 类别中心与距离','',
        '复用 run_006 的同一份34维外观特征：32维经L2归一化的PCA外观特征，加2维稳健标准化、裁剪并乘0.15权重的形状特征。没有重新提取模型特征、重新拟合PCA或加入几何/目录/划分特征。','',
        '每个新类别中心为该类原参考目标的特征算术平均（按目标数自然加权，不把各A簇等权平均）。中心仅使用原有类别归属的6,982个目标计算并冻结；不把本次离群归类加入中心，也不迭代移动中心。每个离群目标按欧氏距离最近的中心归类；并列时取较小Class编号。','',
        '中心数组保存在 `centers.npz`，数值、原参考数量、P95参考半径及代表样本保存在 `centers.json`。特征索引先按原目标身份对齐，避免把缓存位置误作目标编号。','',
        '## 复核依据','',
        f'全部新归类均为机器建议，待人工复核。其中 **{summary["priority_review_targets"]:,} 个优先复核**：{summary["far_from_center_targets"]:,} 个超过分配类别原成员距中心的P95半径；{summary["ambiguous_targets"]:,} 个前两类距离相近；两项有交集，不能直接相加。','',
        '前两类相对间隔定义为 `(第二近距离−最近距离)/第二近距离`，小于0.10标记候选接近；两个距离均为0时记间隔0。参考P95半径按该类原成员距冻结中心的经验95%分位数计算。以上阈值仅用于复核排序，不拒绝分配，也不表示分类概率或正确率。','',
        f'新归类中 {summary["new_domain_targets"]:,} 个目标的配置域未出现在所分配类别的原参考成员中；该信息仅用于解释，未参与距离计算。','',
        f'作为中心代表性的诊断，原参考目标中有 {summary["reference_center_disagreements"]:,} 个的最近中心不等于其原分组，仍保留其原归属。这说明部分合并组分布并非单峰，单一中心无法完全表达组内差异；该诊断不是分类测试精度。','',
        '## 最终类别统计','',*table,'',
        '## 结果字段与追溯','',
        '请使用 `final_class_id` / `final_class` 作为本次最终类别。上一版 `merged_class_id` / `merged_class`、原A类、原聚类编号与离群状态均完整保留，供追溯。`classification_source` 区分原表格合并归属与本次最近中心分配。','',
        '逐目标提供最近与第二近中心、距离、参考P95距离比、前两类间隔、优先复核原因和配置域覆盖情况。`classification_review_required` 对所有新归类为true；优先复核标记只是其中一个子集。原隶属分数仍属于旧HDBSCAN结果，不能当作新类别分数。','',
        '正式标签、原划分、历史合并结果及分割标签不改写；原几何测量和候选结构描述保持不变。数据包含原train/val/test的参考成员，因此本轮是全量数据属性补分组，不构成独立训练/测试实验，也不报告分类测试精度。','',
        '## 复现与核验','',
        '`python -m pallet.reclassify_appearance_outliers`；可指定 `--merged`、`--clustering`、`--output`。输入与输出哈希、冻结来源、中心配置见运行清单和校验和。核验包括5,362个离群目标唯一分配、6,982个原归属保留、全量目标身份和原字段未变、独立距离重算以及正式标签/冻结元数据哈希。','',
        '## 类别参考与新增样本','',
        '每类展示最接近中心的原参考、新归类近中心样本、远中心样本和前两类接近样本，供人工比较。']
    markdown='\n'.join(lines)+'\n';protected_write(ctx['output']/'report_zh.md',markdown.encode())
    body=MarkdownIt('commonmark').enable('table').render(markdown).replace('<table>','<div class="table"><table>').replace('</table>','</table></div>')
    media=relative(ctx['clustering'].parent.parent/'media',ctx['output']);examples=[];panels=''
    for center in centers:
        class_id=center['class_id'];own=[r for r in rows if r['final_class_id']==class_id]
        reference=sorted([r for r in own if not r['original_appearance_outlier']],key=lambda r:(r['assigned_center_distance'],r['target_id']))
        added=[r for r in own if r['original_appearance_outlier']]
        picks=[('原参考近中心',r) for r in reference[:2]]
        picks += [('新增近中心',r) for r in sorted(added,key=lambda r:(r['assigned_center_distance'],r['target_id']))[:2]]
        picks += [('新增远中心',r) for r in sorted(added,key=lambda r:(-r['assigned_center_distance'],r['target_id']))[:2]]
        picks += [('前两类接近',r) for r in sorted(added,key=lambda r:(r['centroid_top2_margin'],r['target_id']))[:2]]
        panels+=f'<section class="group"><h2>{html.escape(center["class_name"])}</h2><p><a href="gallery.html?class={class_id}">查看本类新分配目标</a></p><div class="examples">'
        seen=set()
        for kind,row in picks:
            if row['target_id'] in seen:continue
            seen.add(row['target_id']);examples.append({'class_id':class_id,'sample_kind':kind,'target_id':row['target_id'],'target_key':row['target_key']})
            panels+=f'<a class="example" href="gallery.html?target={row["target_key"]}"><img loading="lazy" src="{html.escape(media,quote=True)}/{row["target_key"]}.jpg" alt="{html.escape(kind)}"><b>{html.escape(kind)}</b><small>距离 {row["assigned_center_distance"]:.3f} · 间隔 {row["centroid_top2_margin"]:.3f}</small><small>{html.escape(row["target_key"])}</small></a>'
        panels+='</div></section>'
    protected_write(ctx['output']/'sample_examples.json',encoded(examples))
    document=f'<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>22类中心离群目标重新分类报告</title><style>{STYLE}</style><main>{body}{panels}</main></html>'
    protected_write(ctx['output']/'report_zh.html',document.encode())
    gallery(ctx,rows,groups,centers)


def verify_inputs(ctx):
    for path,expected in ctx['inputs'].items():
        if digest(path)!=expected:raise ValueError(f'Frozen reclassification input changed: {path}')
    for name,expected in ctx['formal']['protected_files'].items():
        if digest(safe_path(ctx['dataset'],name))!=expected:raise ValueError(f'Formal dataset changed: {name}')


def run(args):
    merged,clustering,output=args.merged.resolve(),args.clustering.resolve(),args.output.resolve()
    formal_path=clustering.parent.parent/'run_manifest.json';formal=read_json(formal_path)
    dataset=Path(formal['binding']['dataset']);font=Path(__file__).parent/'assets/AppearanceReclassificationSans.woff'
    for source in (merged,clustering,dataset):
        if output.is_relative_to(source) or source.is_relative_to(output):raise ValueError('Output must be separate from frozen inputs')
    previous=read_json(merged/'output_manifest.json')
    for name,expected in previous['files'].items():
        if digest(safe_path(merged,name))!=expected:raise ValueError(f'Previous merge deliverable changed: {name}')
    inputs={str(p):digest(p) for p in [Path(__file__).resolve(),font.resolve(),formal_path,
        *(p for p in merged.rglob('*') if p.is_file()),clustering/'features.npz',clustering/'feature_transforms.json',
        clustering/'targets.jsonl',clustering/'summary.json']}
    binding={'kind':'frozen_merged_appearance_centroid_assignment','version':1,'merged':str(merged),
        'clustering':str(clustering),'inputs':inputs,'center':'sample-weighted arithmetic mean of original assigned members',
        'feature_space':'run_006 cached 34-dimensional appearance_features, unchanged',
        'distance':'euclidean','radius_quantile':.95,'small_margin_threshold':.10,
        'tie_policy':'smaller Class number','reference_assignments_retained':True,
        'centers_updated_with_outliers':False,'labels_modified':False}
    output.mkdir(parents=True,exist_ok=True);protected_write(output/'run_manifest.json',encoded(binding))
    ctx={'output':output,'merged':merged,'clustering':clustering,'formal':formal,'dataset':dataset,'inputs':inputs}
    source_rows=read_jsonl(clustering/'targets.jsonl');rows=read_jsonl(merged/'targets.jsonl');groups=read_json(merged/'groups.json')
    if len({r['target_id'] for r in rows})!=len(rows):raise ValueError('Duplicate merged target identity')
    with np.load(clustering/'features.npz',allow_pickle=False) as archive:
        features=align_features(rows,source_rows,archive['appearance_indices'],archive['appearance_features'])
    if features.shape[1]!=34:raise ValueError('Expected the original 34-dimensional appearance feature space')
    labels=np.array([r['merged_class_id'] if r['merged_class_id'] is not None else -1 for r in rows],dtype=np.int64)
    if any((r['appearance_cluster']==-1)!=(labels[i]==-1) for i,r in enumerate(rows)):
        raise ValueError('Only actual appearance outliers may be newly classified')
    result=classify_features(features,labels);group_by_id={g['merged_class_id']:g for g in groups}
    if set(result['class_ids'])!=set(group_by_id):raise ValueError('Reference class coverage mismatch')
    output_rows=enrich_rows(rows,result,group_by_id);final_groups=aggregate(groups,output_rows)
    centers=[]
    for j,class_id in enumerate(result['class_ids']):
        member_indices=np.flatnonzero(labels==class_id)
        nearest=min(member_indices,key=lambda i:(result['distances'][i,j],rows[i]['target_id']))
        centers.append({'class_id':int(class_id),'class_name':group_by_id[int(class_id)]['merged_class'],
            'reference_targets':int(result['reference_counts'][j]),'reference_distance_p95':float(result['radii'][j]),
            'center':result['centers'][j].tolist(),'representative_target_id':rows[nearest]['target_id'],
            'representative_target_key':rows[nearest]['target_key']})
    buffer=io.BytesIO();np.savez_compressed(buffer,class_ids=result['class_ids'],centers=result['centers'],
        reference_counts=result['reference_counts'],reference_distance_p95=result['radii'])
    protected_write(output/'centers.npz',buffer.getvalue());protected_write(output/'centers.json',encoded(centers))
    outliers=[r for r in output_rows if r['original_appearance_outlier']]
    priority=[r for r in outliers if r['classification_priority_review']]
    summary={'total_targets':len(rows),'classes':len(groups),'retained_reference_targets':int((labels!=-1).sum()),
        'reassigned_outliers':len(outliers),'priority_review_targets':len(priority),
        'far_from_center_targets':sum('beyond_reference_p95' in r['classification_review_flags'] for r in outliers),
        'ambiguous_targets':sum('small_top2_margin' in r['classification_review_flags'] for r in outliers),
        'new_domain_targets':sum(not r['assigned_class_domain_supported'] for r in outliers),
        'reference_center_disagreements':sum(r['reference_center_disagreement'] for r in output_rows),
        'remaining_unassigned':0,'labels_modified':False,'new_assignments_human_reviewed':False}
    for name,data in [('targets',output_rows),('outlier_assignments',outliers),('priority_review',priority)]:
        protected_write(output/(name+'.jsonl'),b''.join(encoded(r) for r in data))
        fields=['target_id','target_key','final_class_id','final_class','classification_source','split','domain','group_id',
            'assigned_center_distance','second_center_class_id','second_center_distance','centroid_top2_margin',
            'distance_to_reference_p95_ratio','classification_priority_review','classification_review_flags',
            'assigned_class_domain_supported','original_appearance_class','appearance_cluster','bbox']
        write_csv(output/(name+'.csv'),data if name=='targets' else [{k:r[k] for k in fields} for r in data])
    flat=[]
    for g in final_groups:
        flat.append({k:g[k] for k in ['merged_class_id','merged_class','original_appearance_classes',
            'reference_targets','added_outlier_targets','total_targets','priority_review_targets',
            'far_from_center_targets','ambiguous_targets','new_domain_targets']}|
            {s:g['final_splits'][s] for s in ('train','val','test')})
    write_csv(output/'groups.csv',flat);protected_write(output/'groups.json',encoded(final_groups));protected_write(output/'summary.json',encoded(summary))
    statistics=[['新类别','原A类','原参考数量','新增离群数量','最终数量','优先复核','超P95距离','前两类接近','新配置域','train','val','test']]
    statistics += [[g['merged_class'],', '.join(g['original_appearance_classes']),g['reference_targets'],g['added_outlier_targets'],
        g['total_targets'],g['priority_review_targets'],g['far_from_center_targets'],g['ambiguous_targets'],g['new_domain_targets'],
        *[g['final_splits'][s] for s in ('train','val','test')]] for g in final_groups]
    excel_outliers=[['目标ID','最终类别','最近距离','第二候选类别','第二距离','前两类间隔','相对P95距离','优先复核','复核原因','原划分','配置域']]
    excel_outliers += [[r['target_id'],r['final_class'],r['nearest_center_distance'],f'Class {r["second_center_class_id"]}',
        r['second_center_distance'],r['centroid_top2_margin'],r['distance_to_reference_p95_ratio'],
        '是' if r['classification_priority_review'] else '否',', '.join(r['classification_review_flags']),r['split'],r['domain']] for r in outliers]
    center_sheet=[['新类别','参考目标数','参考P95距离','中心代表目标']]+[[c['class_name'],c['reference_targets'],c['reference_distance_p95'],c['representative_target_id']] for c in centers]
    protected_write(output/'离群目标重新分类结果.xlsx',write_workbook([('最终类别统计',statistics),('离群目标归类',excel_outliers),('冻结类别中心',center_sheet)]))
    protected_write(output/'assets/ReportSans.woff',font.read_bytes())
    protected_write(output/'assets/FONT_LICENSE.txt',(merged/'assets/FONT_LICENSE.txt').read_bytes())
    reports(ctx,output_rows,final_groups,centers,summary)
    # Independently recalculate nearest centers with SciPy and verify all immutable fields.
    from scipy.spatial.distance import cdist
    independent=cdist(features,result['centers'],metric='euclidean')
    if not np.allclose(independent,result['distances'],rtol=1e-12,atol=1e-12):raise ValueError('Independent distance check failed')
    if not np.array_equal(result['class_ids'][np.argmin(independent,axis=1)][labels==-1],result['final_labels'][labels==-1]):
        raise ValueError('Outlier assignment is not the nearest merged center')
    if any(any(r[k]!=v for k,v in old.items()) for old,r in zip(rows,output_rows)):
        raise ValueError('Original target fields changed')
    if any(r['final_class_id']!=r['merged_class_id'] for r in output_rows if not r['original_appearance_outlier']):
        raise ValueError('Existing class membership changed')
    if len({r['target_id'] for r in outliers})!=len(outliers) or sum(g['total_targets'] for g in final_groups)!=len(rows):
        raise ValueError('Missing or duplicate target assignment')
    verify_inputs(ctx)
    validation={'state':'passed','all_targets':len(rows),'assigned_outliers':len(outliers),'centers':len(centers),
        'feature_target_identity_aligned':True,'independent_distances_verified':True,'all_original_fields_unchanged':True,
        'existing_assignments_unchanged':True,'centers_frozen_before_outlier_assignment':True,
        'all_new_assignments_marked_unreviewed':all(r['classification_review_required'] for r in outliers),
        'previous_merge_files_unchanged':len(previous['files']),
        'formal_protected_files_unchanged':len(formal['protected_files'])}
    protected_write(output/'validation.json',encoded(validation))
    files={str(p.relative_to(output)):digest(p) for p in output.rglob('*') if p.is_file() and p.name not in ('output_manifest.json','checksums.sha256')}
    protected_write(output/'output_manifest.json',encoded({'files':files,'summary':summary,'validation':validation}))
    protected_write(output/'checksums.sha256',''.join(f'{sha}  {name}\n' for name,sha in sorted(files.items())).encode())
    print(json.dumps({'output':str(output),**summary,'validation':validation},ensure_ascii=False))


def main(argv=None):
    parser=argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--merged',type=Path,default=DEFAULT_MERGED)
    parser.add_argument('--clustering',type=Path,default=DEFAULT_CLUSTERING)
    parser.add_argument('--output',type=Path,default=DEFAULT_OUTPUT)
    run(parser.parse_args(argv))


if __name__=='__main__':main()
