")
+s=s.replace("$('#detailBody').innerHTML=head+top+note+c.examples.map(v=>sampleBlock(D.samples[v.id],v)).join('');", "const gaps=D.coverage.examples[key]||[],gapIds=new Set(gaps.map(v=>v.id)),other=c.examples.filter(v=>!gapIds.has(v.id));\n $('#detailBody').innerHTML=head+(c.coverage?'
类原型与其最近邻 '+top+note+'':top+note)+(gaps.length?'
本类实际未覆盖成员 '+gaps.map(gapBlock).join(''):'')+(other.length?'
其他同类成员 '+other.map(v=>sampleBlock(D.samples[v.id],v)).join(''):'');")
+s=s.replace("if(two.length)body+=", "if(D.coverage.sources[src])body+=coverageBar(D.coverage.sources[src]);\n if(two.length)body+=")
+s=s.replace("function render(){poses.clear();", "function render(){poses.clear();renderCoverage();")
+s=s.replace("$('#hand').onchange", "$('#coverageFilter').onchange=e=>{coverageFilter=e.target.value;render()};$('#coverageSort').onchange=e=>{coverageSort=e.target.value;render()};\n$('#hand').onchange")
+p.write_text(s)
+p=root/'style.css';s=(Path('/SSD3/yangzijing/Hand_pose_recluster_distinct_20261014/code/web')/p.name).read_text()+'''
+.coverage-panel{background:#fff;border:1px solid var(--line);border-radius:8px;padding:14px 16px;margin-bottom:16px}.coverage-controls{display:flex;align-items:center;gap:12px;flex-wrap:wrap;font-size:13px}.coverage-controls strong{margin-right:auto}.coverage-summary{display:flex;align-items:center;gap:20px;flex-wrap:wrap;margin-top:12px}.coverage-summary .coverage-breakdown{flex:1;min-width:260px}.coverage-bar{height:9px;display:flex;border-radius:5px;overflow:hidden;background:#e7edf2;width:100%;min-width:100px}.covered{background:#218762}.low{background:#c88926}.uncovered{background:#ca5054}.dot{width:9px;height:9px;display:inline-block;border-radius:50%;margin-right:3px}.coverage-values{display:flex;gap:9px;justify-content:space-between;font-size:11px;flex-wrap:wrap;margin-top:4px}.covered-text{color:#146847}.low-text{color:#8e5706}.uncovered-text{color:#a73139}.coverage-legend{font-size:12px;margin:10px 0 4px}.coverage-badge{display:inline-block;padding:2px 7px;border-radius:5px;font-size:11px;white-space:nowrap}.coverage-badge.covered{background:#e3f3eb;color:#146847}.coverage-badge.low{background:#fff1d9;color:#8e5706}.coverage-badge.uncovered{background:#fce7e8;color:#a73139}.card .coverage-breakdown{padding:4px 11px 7px}.card-head>span{display:flex;gap:3px;flex-wrap:wrap;justify-content:flex-end}.coverage-panel details{margin-top:8px;padding:7px 10px;font-size:12px;background:#fafcfe}.coverage-panel .hint{margin:7px 0}.gap-case{border:1px solid #ead4d5;background:#fffafa;padding:16px;border-radius:8px;margin:14px 0}.gap-case h3{font-size:16px;margin:0 0 5px}.gap-case h4{font-size:13px;margin:10px 0 5px}.gap-case details{font-size:12px;padding:7px;margin-top:10px}.gap-case .pair .gap-image{width:100%;height:220px;max-height:none;object-fit:contain;background:#f0f3f6}.sample-coverage{font-size:12px;color:var(--muted);line-height:1.9}.gap-title{color:#a73139;border-top:1px solid var(--line);padding-top:16px}.scroll .coverage-breakdown{min-width:230px}.coverage-breakdown .metric{margin:5px 0}.coverage-panel summary{font-size:12px}.row .coverage-breakdown{max-width:720px;margin-bottom:12px}
+@media(max-width:620px){.coverage-panel{padding:11px}.coverage-controls{gap:8px}.coverage-controls strong{width:100%}.coverage-summary{gap:8px}.card .coverage-breakdown{padding:3px 7px 6px}.card .coverage-values{font-size:9px;gap:1px;display:grid;grid-template-columns:1fr}.card-head{gap:4px;align-items:flex-start}.card-head>span{max-width:80%}.gap-case{padding:9px}.gap-case .pair{grid-template-columns:1fr}.gap-case .pair .gap-image{height:auto;max-height:300px}.coverage-legend{line-height:1.9}.coverage-controls select{max-width:200px}}
+''';p.write_text(s)
+print('UPDATED',root)
diff --git a/pose_recluster_distinct_20261014/downloads/coverage/code/validate_coverage.py b/pose_recluster_distinct_20261014/downloads/coverage/code/validate_coverage.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ceaf63b8ffa3d2e4f6925caba653d4152945557
--- /dev/null
+++ b/pose_recluster_distinct_20261014/downloads/coverage/code/validate_coverage.py
@@ -0,0 +1,35 @@
+from cluster_common import *
+BASE=Path('/SSD3/yangzijing/Hand_pose_recluster_distinct_20261014')
+def main():
+ data=np.load(ROOT/'cache/external_coverage.npz');classes=read(ROOT/'metrics/class_coverage.json');rows=read(BASE/'manifests/external_members.json');ours=read(BASE/'manifests/ours_members.json')
+ ob=np.load(BASE/'cache/ours_features.npz')['bones'];eb=np.load(BASE/'cache/external_features.npz')['bones'];rng=np.random.default_rng(2026101408);checks=[]
+ for side in ['left','right']:
+ bx=np.array([i for i,m in enumerate(ours) if m['hand']==side]);qx=np.array([i for i,m in enumerate(rows) if m['hand']==side]);bank=ob[bx].astype(np.float64)
+ units=[ours[i]['independent_unit'] for i in bx];unames=sorted(set(units));umap={u:i for i,u in enumerate(unames)};uc=np.array([umap[u] for u in units])
+ boundary=qx[np.argsort(np.abs(data['distance_deg'][qx]-data['threshold_deg'][qx]))[:4]]
+ samples=list(map(int,rng.choice(qx,12,replace=False)))+list(map(int,boundary))
+ for i in samples:
+ d=np.degrees(np.arccos(np.clip(np.einsum('njk,jk->nj',bank,eb[i].astype(np.float64)),-1,1))).mean(-1)
+ j=int(np.argmin(d));err=abs(float(d[j])-data['distance_deg'][i]);assert err<.01,(i,err)
+ nn=int(data['nearest_index'][i]);assert abs(float(d[np.where(bx==nn)[0][0]])-float(d[j]))<.01
+ tau=float(data['threshold_deg'][i]);records=len(set(uc[d<=tau]));count=int((d<=tau).sum());status=2 if d[j]>tau else (0 if records>=3 else 1)
+ margin=float(np.min(abs(d-tau)));ambiguous=margin<.002
+ if not ambiguous:assert records==data['support_recordings'][i] and count==data['support_samples'][i]
+ assert status==data['status'][i],(i,'classification mismatch')
+ checks.append(dict(index=i,id=rows[i]['id'],hand=side,max_nn_error_deg=err,cpu_nearest_index=int(bx[j]),gpu_nearest_index=nn,cpu_support_recordings=records,gpu_support_recordings=int(data['support_recordings'][i]),cpu_support_samples=count,gpu_support_samples=int(data['support_samples'][i]),threshold_boundary_margin_deg=margin,numerically_ambiguous_counts=ambiguous,status_agrees=True))
+ # Numerical comparison with existing independently calculated exact class-prototype NN.
+ models=read(BASE/'metrics/models.json');proto=[]
+ for side in ['left','right']:
+ for level in ['fine','distinct']:
+ for c in models['external_'+side][level]:
+ i=c['prototype_index'];err=abs(data['distance_deg'][i]-c['nearest_reference']['mean_deg']);assert err<.01,(c['key'],err)
+ proto.append(float(err))
+ assert sum(c['n'] for c in classes.values() if c['level']=='distinct')==len(rows)
+ assert sum(c['n'] for c in classes.values() if c['level']=='fine')==len(rows)
+ assert all(c['covered']+c['low']+c['uncovered']==c['n'] for c in classes.values())
+ assert all(ours[int(nn)]['hand']==r['hand'] for r,nn in zip(rows,data['nearest_index']))
+ hashes=read(ROOT/'audit/input_sha256.json');assert all(sha(p)==h for p,h in hashes.items()),'Input changed'
+ report=dict(status='PASS',n_queries=len(rows),same_hand_all=True,all_class_counts_valid=True,input_hashes_unchanged=True,cpu_float64_checks=checks,cpu_nn_max_error_deg=max(c['max_nn_error_deg'] for c in checks),previous_exact_prototype_checks=len(proto),previous_exact_prototype_max_error_deg=max(proto),limitations='Float32 acos counts extremely close to the threshold can change under rounding; boundary examples were checked and sample statuses agree. No GT accuracy verdict derived from nearest-neighbor distance.')
+ write(ROOT/'audit/numerical_validation.json',report);print({k:v for k,v in report.items() if k!='cpu_float64_checks'})
+
+if __name__=='__main__':main()
diff --git a/pose_recluster_distinct_20261014/downloads/coverage/code/verify_public_assets.py b/pose_recluster_distinct_20261014/downloads/coverage/code/verify_public_assets.py
new file mode 100644
index 0000000000000000000000000000000000000000..f303ef0acc6d25a42a373745592cbe67947fcd6e
--- /dev/null
+++ b/pose_recluster_distinct_20261014/downloads/coverage/code/verify_public_assets.py
@@ -0,0 +1,38 @@
+"""Check permanent public URLs, bytes and archived previous entry after publication."""
+from cluster_common import *
+import requests,time,concurrent.futures,re
+BASE_URL='https://yangzijing-nine-source-hand-data-browser.static.hf.space/'
+
+def get_verified(item):
+ url=BASE_URL+item['path'];events=[]
+ for attempt in range(4):
+ try:
+ r=requests.get(url,timeout=(10,50));r.raise_for_status();data=r.content
+ if item['path'].endswith('.html'):
+ # Static Spaces adds only this known runtime creator-ID script.
+ data=re.sub(rb'',b'',data)
+ actual=hashlib.sha256(data).hexdigest();assert actual==item['sha256'],'SHA256 mismatch'
+ return dict(path=item['path'],status=200,sha256=actual,bytes=len(data),retries=attempt,events=events)
+ except Exception as e:
+ events.append(dict(attempt=attempt,error=type(e).__name__,http_status=getattr(getattr(e,'response',None),'status_code',None)))
+ if attempt==3:return dict(path=item['path'],error=type(e).__name__+': '+str(e)[:180],events=events)
+ time.sleep(2*(attempt+1))
+
+def main():
+ plan=read(ROOT/'audit/publication_plan.json');receipt=read(ROOT/'audit/publication.json');assert 'space' in receipt
+ entry=next(x for x in plan['files'] if x['relative']=='index.html')
+ for attempt in range(36):
+ r=requests.get(BASE_URL+entry['path'],params={'coverage_revision':receipt['space']['commit']},timeout=(10,40))
+ if r.status_code==200 and 'data.coverage_v2.js' in r.text:break
+ time.sleep(5)
+ else:raise RuntimeError('New static entry not visible after deployment wait')
+ checks=[]
+ with concurrent.futures.ThreadPoolExecutor(2) as pool:
+ fs={pool.submit(get_verified,item):item for item in plan['files']}
+ for f in concurrent.futures.as_completed(fs):
+ checks.append(f.result())
+ if len(checks)%50==0:print('PUBLIC_CHECKED',len(checks),'/',len(fs),flush=True)
+ bad=[r for r in checks if 'error' in r];out=dict(status='PASS' if not bad else 'FAIL',url=receipt['url'],commit=receipt['space']['commit'],files=len(checks),failures=bad,retried=sum(r.get('retries',0)>0 for r in checks),checks=checks)
+ write(ROOT/'audit/public_assets.json',out);print('PUBLIC_ASSETS',out['status'],out['files'],'failed',len(bad),flush=True)
+ assert not bad,str(bad[:3])
+if __name__=='__main__':main()
diff --git a/pose_recluster_distinct_20261014/downloads/coverage/code/visualize.py b/pose_recluster_distinct_20261014/downloads/coverage/code/visualize.py
new file mode 100644
index 0000000000000000000000000000000000000000..2f2fdc338fc9cb388cdb9735683e4a33315a3108
--- /dev/null
+++ b/pose_recluster_distinct_20261014/downloads/coverage/code/visualize.py
@@ -0,0 +1,166 @@
+"""Paired RGB and calibrated GT overlay for members of the newly fitted clusters."""
+from cluster_common import *
+import sys,functools,pickle,zipfile,shutil
+from PIL import Image
+import cv2
+cv2.setNumThreads(1)
+
+sys.path.insert(0,str(REF/'code'))
+import visuals as legacy
+import interhand_rgb
+legacy.ROOT=ROOT;interhand_rgb.ROOT=ROOT
+from geometry import project
+import torch,smplx
+torch.set_num_threads(1);torch.set_grad_enabled(False)
+
+EXTRA=Path('/SSD3/yangzijing/Hand_missing_source_visuals_20261012')
+COLOR=[(60,180,210),(80,170,40),(230,100,40),(180,80,180),(80,80,230)]
+
+@functools.lru_cache(2)
+def mano(side):
+ m=smplx.MANO('/SSD3/jiasen/foundation_hand/pretrained/mano',is_rhand=side=='right',use_pca=False,flat_hand_mean=True).eval()
+ if side=='left' and torch.abs(m.shapedirs[:,0,:]-mano('right').shapedirs[:,0,:]).sum()<1:m.shapedirs[:,0,:]*=-1
+ return m
+@functools.lru_cache(2)
+def arc(path):return zipfile.ZipFile(path)
+def blob(path):
+ if '::' in path:
+ a,b=path.split('::',1);return arc(a).read(b)
+ return Path(path).read_bytes()
+
+def ho_image(m):
+ source=blob(m['annotation_path']);assert hashlib.sha256(source).hexdigest()==m['annotation_sha256']
+ d=pickle.loads(source,encoding='latin1');field='hand' if m['source']=='wilor_ho3d' else m['hand']+'Hand'
+ p=np.asarray(d[field+'Pose']);t=np.asarray(d[field+'Trans']);beta=np.asarray(d['handBeta']);model=mano(m['hand'])
+ tensor=lambda a:torch.as_tensor(np.asarray(a).copy()[None],dtype=torch.float32)
+ out=model(global_orient=tensor(p[:3]),hand_pose=tensor(p[3:]),betas=tensor(beta),transl=tensor(t));v=out.vertices[0].numpy()
+ err=float(np.linalg.norm(out.joints[0,:16].numpy()-np.asarray(d[field+'Joints3D'])[:16],axis=-1).max()*1000);assert err<.02
+ v=v*np.array([1,-1,-1]);K=np.asarray(d['camMat']);rgb=blob(m['image_path']);im=cv2.imdecode(np.frombuffer(rgb,np.uint8),cv2.IMREAD_COLOR);assert im is not None
+ uv=project(v,K);mask=(v[:,2]>0)&(uv[:,0]>=0)&(uv[:,0]
=0)&(uv[:,1]=500,'target mostly out of image'
+ vertices=np.full((2,778,3),np.nan);side=['left','right'].index(m['hand']);vertices[side]=v;valid=[False,False];valid[side]=True
+ overlay,pixels=legacy.renderer.media.render_overlay(im,vertices,valid,lambda pts:project(pts,K));box,inside=legacy.renderer.bbox(uv,im.shape[:2]);assert inside
+ raw=Image.fromarray(im[...,::-1]);ov=Image.fromarray(overlay[...,::-1]);full=Image.new('RGB',(raw.width*2,raw.height));full.paste(raw);full.paste(ov,(raw.width,0))
+ ev=dict(camera='same-source unmodified camMat; OpenGL->OpenCV diag(1,-1,-1); original full image',K=K.tolist(),source_rgb_sha256=hashlib.sha256(rgb).hexdigest(),annotation_sha256=m['annotation_sha256'],source_reconstruction_max_mm=err,independent_3d_reference=False,crop_native_xyxy=list(box),target_in_frame=True,native_hw=list(im.shape[:2]),mesh_pixels=pixels)
+ return full,raw.crop(box),ov.crop(box),ev
+
+@functools.lru_cache(10)
+def reinter_joints(path):
+ z=np.load(path);return {int(f):j for f,j in zip(z['frames'],z['joints_mm'])}
+def joint_pair(im,uv,mask,evidence):
+ ov=im.copy();mask=mask & np.isfinite(uv).all(-1)&(uv[:,0]>=0)&(uv[:,0]=0)&(uv[:,1]=12,'too few target joints in frame'
+ for k in range(1,21):
+ parent=int(PARENTS[k]);color=COLOR[(k-1)//4]
+ if mask[k] and mask[parent]:cv2.line(ov,tuple(np.rint(uv[parent]).astype(int)),tuple(np.rint(uv[k]).astype(int)),color,max(2,round(im.shape[0]/500)),cv2.LINE_AA)
+ for k in np.flatnonzero(mask):cv2.circle(ov,tuple(np.rint(uv[k]).astype(int)),max(2,round(im.shape[0]/500)),(255,255,255),-1,cv2.LINE_AA)
+ box,ok=legacy.renderer.bbox(uv[mask],im.shape[:2]);assert ok
+ raw=Image.fromarray(im[...,::-1]);over=Image.fromarray(ov[...,::-1]);full=Image.new('RGB',(raw.width*2,raw.height));full.paste(raw);full.paste(over,(raw.width,0))
+ evidence.update(crop_native_xyxy=list(box),target_in_frame=True,native_hw=list(im.shape[:2]),joints_drawn=int(mask.sum()))
+ return full,raw.crop(box),over.crop(box),evidence
+
+@functools.lru_cache(1)
+def reinter_index():return read(ROOT/'manifests/reinter_rgb.json')
+def reinter_image(m):
+ for item in reinter_index().get(m['id'],[]):
+ camera=read(item['camera_file'])[item['camera']];j=reinter_joints(m['annotation_path'])[m['frame']][['left','right'].index(m['hand']),m['source_joint_mapping_24_to_21']]
+ xyz=j@np.asarray(camera['R']).T+np.asarray(camera['t']);f=np.asarray(camera['focal']);pp=np.asarray(camera['princpt']);uv=xyz[:,:2]/xyz[:,2:]*f+pp;K=np.array([[f[0],0,pp[0]],[0,f[1],pp[1]],[0,0,1.]])
+ check=cv2.projectPoints(xyz,np.zeros(3),np.zeros(3),K,np.zeros(5))[0][:,0];assert abs(check-uv).max()<1e-5
+ im=cv2.imread(item['path']);assert im is not None and sha(item['path'])==item['image_sha256']
+ e=dict(camera='official Mugsy camera R*X+t, native image pixels, no additional distortion by source renderer',K=K.tolist(),camera_id=item['camera'],camera_sha256=item['camera_sha256'],source_rgb_sha256=item['image_sha256'],source_rgb_member=item['member'],annotation_sha256=m['annotation_sha256'],projection_crosscheck_max_px=float(abs(check-uv).max()),label='official original-fit renderer3D joints; no MANO fit',joint_visibility_independent=False)
+ try:return joint_pair(im,uv,xyz[:,2]>0,e)
+ except AssertionError:continue
+ raise ValueError('No readable selected ReInterHand view')
+
+@functools.lru_cache(1)
+def ih_index():return read(ROOT/'manifests/interhand_rgb.json')
+@functools.lru_cache(1)
+def existing():
+ result={}
+ for p in (REF/'audit/media').glob('*.json'):
+ if p.name.endswith('_acquisition.json'):continue
+ r=read(p)
+ if r.get('assets') and all((REF/'site'/v).exists() for v in r['assets'].values()):result[r['id']]=r
+ return result
+
+def paired(m):
+ if m['id'] in existing():return True
+ if m['source'] in ['wilor_rhd','wilor_freihand']:return local_image(m) is not None
+ if m['source']=='wilor_interhand26m':return m['id'] in ih_index()
+ if m['source']=='wilor_reinter':return m['id'] in reinter_index()
+ if m['source']=='wilor_mtc':
+ c=legacy.extras().get((m['source'],str(m.get('subject')),int(m['frame']),m['hand']))
+ return c is not None and c['recording']==m['sequence']
+ return True
+
+def local_image(m):
+ src=m['source'];rel=f"training/rgb/{m['frame']:08d}.jpg" if src=='wilor_freihand' else f"RHD_published_v2/training/color/{m['frame']:05d}.png"
+ paths=[root/'sources'/src/rel for root in [ROOT,Path('/SSD3/yangzijing/Hand_pose_recluster_distinct_20261014'),REF,OLD]]
+ paths.append(legacy.OLD/'sources'/('FREIHAND' if src=='wilor_freihand' else 'RHD')/rel)
+ return next((p for p in paths if p.exists()),None)
+
+def save(m):
+ p=ROOT/'audit/media'/f"{m['id']}.json"
+ if p.exists():return read(p)
+ if m['id'] in existing():
+ old=existing()[m['id']];r=dict(old);r['reused_from']=str(REF);r['assets']={}
+ for kind,rel in old['assets'].items():
+ dst=ROOT/'site/media'/Path(rel).name;shutil.copyfile(REF/'site'/rel,dst);r['assets'][kind]='media/'+dst.name
+ write(p,r);return r
+ if m['source'] in ['wilor_ho3d','wilor_h2o3d']:full,raw,ov,e=ho_image(m)
+ elif m['source']=='wilor_reinter':full,raw,ov,e=reinter_image(m)
+ elif m['source'] in ['wilor_rhd','wilor_freihand']:
+ mm=dict(m);mm['image_path']=str(local_image(m));assert local_image(m) is not None
+ if m['source']=='wilor_freihand':
+ K,t,j=legacy.fh_arrays();_,jj,_=legacy.freihand_decode(t[m['frame']],K[m['frame']]);mm['native_joint_mean_error_mm']=float(np.linalg.norm(jj[0]-j[m['frame']],axis=-1).mean()*1000)
+ full,raw,ov,e=legacy.renderer.native(mm)
+ else:full,raw,ov,e=legacy.image_for(m)
+ assert e.get('target_in_frame') and raw.size==ov.size and min(raw.size)>=24,'target not readable'
+ change=np.abs(np.asarray(raw,dtype=float)-np.asarray(ov,dtype=float)).max(-1)>14;assert change.sum()>=12,'overlay absent at target'
+ nonblack=float((np.asarray(raw).max(-1)[change]>15).mean());assert nonblack>=.55,'projection mainly black padding'
+ if 'in_frame_vertices_per_hand' in e:assert e['in_frame_vertices_per_hand'][['left','right'].index(m['hand'])]>=500,'target mostly out of image'
+ pair=Image.new('RGB',(raw.width*2,raw.height));pair.paste(raw);pair.paste(ov,(raw.width,0));assets={}
+ for kind,im in [('full',full),('crop',pair)]:
+ dest=ROOT/'site/media'/f"{m['id']}_{kind}.jpg";im.save(dest,quality=90);assets[kind]='media/'+dest.name
+ label='EgoRenderer Ver.005 既有拟合 MANO' if m['source']=='synthetic' else ('来源 GT 关节投影' if m['source'] in ['wilor_reinter','wilor_interhand26m','wilor_mtc','wilor_rhd'] else '来源拟合 MANO')
+ r=dict(id=m['id'],assets=assets,evidence=e,label=label,gt_status='证据不足',legibility=dict(changed_pixels=int(change.sum()),non_black_GT_pixel_fraction=nonblack),sha256={v:sha(ROOT/'site'/v) for v in assets.values()})
+ write(p,r);return r
+
+def setup():
+ p=ROOT/'manifests/interhand_rgb.json'
+ if not p.exists():
+ data=read(REF/'manifests/interhand_rgb.json')
+ (ROOT/'sources/interhand_rgb').mkdir(exist_ok=True)
+ for sid,x in data.items():
+ if not Path(x['path']).exists():x['path']=str(ROOT/'sources/interhand_rgb'/f'{sid}.jpg')
+ write(p,data)
+
+def main(group,level_arg=None):
+ setup();rows,b=load(group);result={};failures=[];failed=set();dest=ROOT/'metrics'/f'{group}_visuals.json'
+ exclusions=ROOT/'audit/visual_review_exclusions.json'
+ if exclusions.exists():failed.update(read(exclusions))
+ preference_path=ROOT/'manifests/visual_review_preferences.json'
+ preferences=read(preference_path) if preference_path.exists() else {}
+ if dest.exists():result=read(dest)
+ for side in ['left','right']:
+ model=read(ROOT/'metrics'/f'{group}_{side}.json');a=np.load(ROOT/'cache'/f'{group}_{side}_assignments.npz');ix=a['indices']
+ for level in ([level_arg] if level_arg else ['main','fine']):
+ label=np.load(ROOT/'cache'/f'{group}_{side}_distinct_assignments.npz')['distinct'] if level=='distinct' else a[level]
+ for c in model[level]:
+ if c['key'] in result:continue
+ ids=ix[label==c['label']];distance=angular(b[[c['prototype_index']]],b[ids])[0];order=np.argsort(distance,kind='stable');picked=[];units=set()
+ if c['key'] in preferences:
+ preferred=preferences[c['key']]['sample_ids'];positions={rows[int(i)]['id']:pos for pos,i in enumerate(ids)}
+ assert all(sid in positions for sid in preferred),'Visual preference must stay in the actual GT class'
+ lead=[positions[sid] for sid in preferred];order=np.array(lead+[int(p) for p in order if int(p) not in lead])
+ for pos in order:
+ i=int(ids[pos]);m=rows[i]
+ if m['id'] in failed or m['independent_unit'] in units or not paired(m):continue
+ try:r=save(m)
+ except Exception as e:
+ failures.append(dict(key=c['key'],id=m['id'],source=m['source'],error=type(e).__name__+': '+str(e)[:220]));failed.add(m['id']);write(ROOT/'audit'/f'{group}_media_failures.json',failures);continue
+ picked.append(dict(index=i,id=m['id'],source=m['source'],recording=m['recording'],frame=m['frame'],hand=m['hand'],label_type=m['label_type'],prototype_distance_deg=float(distance[pos]),is_prototype=i==c['prototype_index'],joints=standard(b[i]),media=r));units.add(m['independent_unit'])
+ if len(picked)>=2:break
+ assert picked,(c['key'],'No same-cluster paired image')
+ result[c['key']]=picked;write(dest,result);print('VISUAL',c['key'],[(p['source'],round(p['prototype_distance_deg'],2)) for p in picked],flush=True)
+ write(ROOT/'audit'/f'{group}_media_failures.json',failures)
+
+if __name__=='__main__':main(*sys.argv[1:])
diff --git a/pose_recluster_distinct_20261014/downloads/coverage/code/web/app.js b/pose_recluster_distinct_20261014/downloads/coverage/code/web/app.js
new file mode 100644
index 0000000000000000000000000000000000000000..7025f1d9b236fa41f6e5fc155dbe2430002b5eb0
--- /dev/null
+++ b/pose_recluster_distinct_20261014/downloads/coverage/code/web/app.js
@@ -0,0 +1,61 @@
+const D=window.AUDIT_DATA;
+// Retry transient failures on the image CDN while keeping permanent source URLs.
+document.addEventListener('error',e=>{const im=e.target;if(im?.tagName!=='IMG')return;const n=Number(im.dataset.networkRetry||0);if(n<2){im.dataset.networkRetry=String(n+1);setTimeout(()=>{if(im.isConnected)im.src=im.getAttribute('src')},600*(n+1))}},true);
+const $=s=>document.querySelector(s), nf=n=>Number(n).toLocaleString('en-US'), esc=s=>String(s??'').replace(/[&<>"']/g,c=>({'&':'&','<':'<','>':'>','"':'"',"'":'''}[c]));
+let hand='right',level='distinct',view='oblique',tab='overview',poseSerial=0,activeDetail='',coverageFilter='all',coverageSort='gap';
+const colors=['#b8852f','#39896e','#437eb0','#9061ab','#c66e70'];
+const poses=new Map(), allClasses={};
+Object.values(D.models).forEach(m=>['fine','distinct'].forEach(l=>m[l].forEach(c=>allClasses[c.key]=c)));
+const model=g=>D.models[g+'_'+hand], cname=s=>D.names[s]||s;
+function poseCanvas(joints,label='',overlay=null){const id='pose'+poseSerial++;poses.set(id,{joints,overlay,rot:[0,0],label});return `
已覆盖 ${pct(v.covered_fraction)} 低覆盖 ${pct(v.low_fraction)} 未覆盖 ${pct(v.uncovered_fraction)}
${compact?'':`
分母 ${nf(v.n)} 个实际分析手帧：${nf(v.covered)} 已覆盖 / ${nf(v.low)} 低覆盖 / ${nf(v.uncovered)} 未覆盖。
`}
${label} `}
+function filteredExternal(){return [...model('external')[level]].filter(c=>coverageFilter==='all'||coverageFilter==='entire'&&c.coverage.uncovered===c.n||coverageFilter==='mostly_gap'&&c.coverage.uncovered_fraction>=.5||coverageFilter==='mostly_covered'&&c.coverage.covered_fraction>=.7).sort((a,b)=>coverageSort==='id'?a.number-b.number:coverageSort==='covered'?b.coverage.covered_fraction-a.coverage.covered_fraction:b.coverage.uncovered_fraction-a.coverage.uncovered_fraction)}
+function renderCoverage(){const v=D.coverage.summary[hand],cal=D.coverage.calibration[hand],cs=model('external')[level],entire=cs.filter(c=>c.coverage.status==='uncovered').length;$('#coverageSummary').innerHTML=`${coverageBar(v,true)}${hand==='left'?'左':'右'}手 ${nf(v.n)} 个外部手帧 · ${cs.length} 类中 ${entire} 类整类未覆盖 · 相似阈值 ${cal.thresholds_deg.q95.toFixed(2)}°
`;
+ const sens=D.coverage.sensitivity.filter(x=>x.key==='external_'+hand&&['q90','q95','q99'].includes(x.threshold));$('#coverageMethodBody').innerHTML=`距离使用来源 GT 的 20 根 3D 骨方向，在掌部坐标系中计算平均夹角，去除手腕整体运动和骨长差异；左右手分开。不是 2D 图上距离，也不是 MANO 参数直接相减。
参照：H2O＋ARCTIC＋HOT3D＋EgoRenderer Ver.005，共 271,543 手帧；本侧 ${nf(cal.reference_n)} 手帧、${nf(cal.reference_units)} 段独立录制/源动作。用本侧 4,096 个内部留出样本校准，检索时排除同录制全部帧；四源及各录制均衡加权，取最近距离的 95% 分位。每个外部成员穷举检索全部同侧参照。
校准分位 阈值 已覆盖 低覆盖 未覆盖 ${sens.map(x=>`${x.threshold==='q95'?'95%（本页）':x.threshold.slice(1)+'%'} ${x.threshold_deg.toFixed(2)}° ${pct(x.covered_fraction)} ${pct(x.low_fraction)} ${pct(x.uncovered_fraction)}
类标签：≥90% 成员已覆盖才标“≥90% 已覆盖”；所有成员无近邻才标“整类未覆盖”；其余标“部分覆盖”。90% 是显示约定，完整比例始终保留。统计分母是本轮分析样本，不是 WiLoR 官方全量；2D 来源不作 3D 覆盖判断。我们参照使用已核对训练成员的来源 GT，最终 canonical 标签和 loss mask 尚未验收。
`}
+function memberBadge(s){const v=s.coverage;if(!v)return '';return `${{covered:'已覆盖',low:'低覆盖',uncovered:'抽样库未覆盖'}[v.status]} 最近距离 ${v.distance_deg.toFixed(2)}° / 阈值 ${v.threshold_deg.toFixed(2)}° · ${nf(v.support_recordings)} 段相似参照录制
`}
+function gapBlock(v){const s=D.samples[v.id],n=v.nearest;return `未覆盖成员 · ${esc(cname(s.source))} · ${s.hand==='left'?'左手':'右手'} · 帧 ${s.frame} ${esc(s.recording)} · ${esc(s.label)} · 距该类原型 ${v.prototype_distance_deg.toFixed(2)}°
${memberBadge(s)}外部实际成员 ${poseCanvas(v.joints,'该未覆盖成员的 GT 姿态')}
原图 ｜ GT 叠加 · 点击看完整原画面
我们库中的精确最近邻 ${poseCanvas(n.joints,'该成员的精确最近 GT 姿态')}${n.media?`
原图 ｜ GT 叠加 · 点击看完整原画面
`:'
这个精确最近邻未通过原图可读性检查，保留其 GT 骨架；没有用其他样本替代。
'}
${esc(cname(n.source))} · ${esc(n.recording)} · 帧 ${n.frame} · ${n.hand==='left'?'左':'右'}手
相机与 GT 证据 外部：${esc(s.camera)}。${esc(s.gt_status)}
最近邻：${esc(n.media?.evidence?.camera??n.media_missing_reason)}。图像叠加仅供核查，覆盖距离不构成 GT 可靠性结论。
${c.id} ${classBadge(c)}${related(c)&&level==='distinct'?'有相近类 ':''} ${nf(c.n)} 手帧 · ${Object.keys(c.sources).length} 来源
${c.coverage?coverageBar(c.coverage,true):''}${poseCanvas(c.joints,'GT 类原型')}${c.coverage?'未覆盖成员 / 原图 / 近邻':'原图 / GT 叠加 / 近邻'} ${g==='ours'?'我们三源＋合成':'WiLoR 外部 3D 来源'} ${nf(m.n)} ${m[level].length} ${m.fine.length-m.distinct.length} 组别 · ${hand==='left'?'左手':'右手'} 参与手帧 当前类数 合并减少细类数 ${rows}
`}
+function renderOverview(){countTable();for(const g of ['ours','external']){const cs=g==='external'?filteredExternal():model(g)[level];$('#'+g+'Cards').innerHTML=cs.map(card).join('')||'当前筛选没有类别。
';$('#'+g+'Count').textContent=`${cs.length} / ${model(g)[level].length} 类`}scatter()}
+function pairTile(c){const s=D.samples[c.examples[0].id];return `${poseCanvas(c.joints,'GT 类原型')}
图例 ${esc(cname(s.source))} · 距原型 ${c.examples[0].prototype_distance_deg.toFixed(1)}°
我们 ${a.id} ↔ 外部 ${c.id} ${classBadge(c)}
${coverageBar(c.coverage)}
${pairTile(a)}${pairTile(c)}
类原型差 ${n.class_prototype_mean_deg.toFixed(1)}° · 外部原型最近参照 ${n.mean_deg.toFixed(1)}°（原型示意；覆盖统计来自全类成员）
看该类未覆盖成员、原图与精确最近邻 当前筛选没有类别。
'}
+function renderSimilar(){const summaries=D.summary.filter(s=>s.hand===hand);$('#duplicateSummary').innerHTML=`组别 32 细类中的相近对 合并后相近对 主视图类数 ${summaries.map(s=>`${s.group==='ours'?'我们':'外部'} ${s.stages.fine.near_pairs} ${s.stages.distinct.near_pairs} ${s.stages.distinct.k}
“相近”按原型或数值中心的 20 骨平均夹角 ≤12°，且每根手指平均夹角 ≤20° 判断；这是本页查看粒度，不是通用手势标准。
`;
+ const pp=D.pairs.filter(p=>p.hand===hand&&p.related);$('#relatedPairs').innerHTML=pp.map(p=>{const a=allClasses[`${p.group}_${hand}_distinct_${p.a}`],b=allClasses[`${p.group}_${hand}_distinct_${p.b}`];return `${p.group==='ours'?'我们':'外部'} ${p.a} / ${p.b} ${p.finger_deg.map((v,i)=>`${['拇','食','中','环','小'][i]}指 ${v.toFixed(1)}°`).join(' · ')}
${pairTile(a)}${pairTile(b)}
原型整体 ${p.mean_deg.toFixed(1)}° · 跨录制近邻更靠近另一类的比例 ${(100*p.member_diagnostic.cross_cluster_closer_fraction).toFixed(1)}%
本侧没有满足判据的剩余相近类对。
';
+ $('#mergeRows').innerHTML=D.merges.filter(m=>m.hand===hand).map(m=>`${m.group==='ours'?'我们':'外部'}：${m.children.join(' + ')} → 主类 ${m.cluster} ${nf(m.n)} 手帧
${m.child_keys.map(k=>{const c=allClasses[k];return `
`}).join('')}
来源组 数据集 / 可视化 实际参与手帧 版本与范围 本轮用途 对参照库覆盖 '+D.sources.map(s=>`${s.group} ${esc(s.name)} ${nf(s.cluster_n||0)} ${esc(s.scope)} ${esc(s.role)} ${D.coverage.sources[s.id]?coverageBar(D.coverage.sources[s.id],true):s.group==='我们'?'参照库':s.gallery_source?'与参照重叠，不单计':'不作 3D 判断'} '}
+function sampleBlock(s,v){return `${memberBadge(s)}${esc(cname(s.source))} · ${esc(s.hand)} · 帧 ${esc(s.frame??'—')} ${esc(s.recording??s.split??'')} · ${esc(s.label)}${v?` · 距类原型 ${v.prototype_distance_deg.toFixed(2)}°`:''}
${esc(s.camera??s.note??'')} · ${esc(s.gt_status??'仅可核查相应标注，不参与3D聚类')}
`}
+function openClass(key){const c=allClasses[key],g=key.split('_')[0];let head=`${g==='ours'?'我们':'外部'} · ${key.includes('_left_')?'左手':'右手'} · ${c.id}${key.includes('_fine_')?'（细类）':''} 原型：${esc(cname(c.prototype.source))} · ${esc(c.prototype.recording)} · 帧 ${c.prototype.frame}。${nf(c.n)} 个成员。
`;
+ head+=c.coverage?classBadge(c)+coverageBar(c.coverage):'';
+ let top=`${poseCanvas(c.joints,'精确 GT 类原型')}
统一掌部坐标与骨长，仅用于姿态比较。
`;
+ if(c.nearest_reference){const n=c.nearest_reference;top+=`
${poseCanvas(n.nearest_joints,'全部参照库中的最近 GT 样本')}
对全部同侧三源＋合成样本检索：${n.mean_deg.toFixed(2)}°。不使用散点图距离。
`}else top+=`
来源组成 ${Object.entries(c.sources).sort((a,b)=>b[1]-a[1]).map(([s,n])=>`
${esc(cname(s))}：${nf(n)}
`).join('')}
`;
+ top+='
当前可读 RGB 图例与类原型仍有较大差别。上方骨架才是精确原型；下面图片保留实际成员身份与距离。
':'';
+ const gaps=D.coverage.examples[key]||[],gapIds=new Set(gaps.map(v=>v.id)),other=c.examples.filter(v=>!gapIds.has(v.id));
+ $('#detailBody').innerHTML=head+(c.coverage?'类原型与其最近邻 '+top+note+'本类实际未覆盖成员 '+gaps.map(gapBlock).join(''):'')+(other.length?'其他同类成员 '+other.map(v=>sampleBlock(D.samples[v.id],v)).join(''):'');if(!$('#detail').open)$('#detail').showModal();wirePoses();activeDetail='class='+encodeURIComponent(key);location.hash=activeDetail}
+function openSource(src){const s=D.sources.find(x=>x.id===src),target=s.gallery_source||src,ids=D.gallery[target]||[],two=D.two_d[src]||[];let body=`${esc(s.name)} · ${esc(s.group)} ${esc(s.scope)}
`;
+ if(D.coverage.sources[src])body+=coverageBar(D.coverage.sources[src]);
+ if(two.length)body+=two.map(v=>sampleBlock(v,null)).join('');
+ else if(ids.length){body+='以下图例均属于本轮冻结成员；点击原图可放大。手侧保留逐样本标记。
';body+=ids.slice(0,20).map(id=>sampleBlock(D.samples[id],null)).join('')}
+ else body+=`${src==='wilor_bedlam'?'按用户要求跳过，本轮没有读取或聚类。':'当前聚类使用该源 GT；本页未选到该源配对图例。'}
`;
+ $('#detailBody').innerHTML=body;if(!$('#detail').open)$('#detail').showModal();activeDetail='source='+encodeURIComponent(src);location.hash=activeDetail}
+function scatter(){const c=$('#scatter'),ctx=c.getContext('2d');ctx.clearRect(0,0,c.width,c.height);const all=['ours','external'].flatMap(g=>model(g)[level].map(x=>({...x,g})));let xx=all.map(c=>c.display_xy[0]),yy=all.map(c=>c.display_xy[1]),xmin=Math.min(...xx),xmax=Math.max(...xx),ymin=Math.min(...yy),ymax=Math.max(...yy);const points=[];for(const p of all){let x=45+(p.display_xy[0]-xmin)/(xmax-xmin)*1000,y=400-(p.display_xy[1]-ymin)/(ymax-ymin)*350;ctx.fillStyle=p.g==='ours'?'#276a9c':'#79569d';ctx.beginPath();ctx.arc(x,y,6,0,Math.PI*2);ctx.fill();ctx.fillText(p.id,x+8,y+4);points.push({x,y,key:p.key})}c.onclick=e=>{const r=c.getBoundingClientRect(),x=(e.clientX-r.left)*c.width/r.width,y=(e.clientY-r.top)*c.height/r.height;const hit=points.find(p=>Math.hypot(x-p.x,y-p.y)<15);if(hit)openClass(hit.key)}}
+function render(){poses.clear();renderCoverage();if(tab==='overview')renderOverview();if(tab==='compare')renderCompare();if(tab==='similar')renderSimilar();if(tab==='sources')renderSources();wirePoses()}
+function changeTab(t){tab=t;document.querySelectorAll('.tab').forEach(e=>e.hidden=e.id!==t);document.querySelectorAll('nav button').forEach(e=>e.classList.toggle('active',e.dataset.tab===t));render()}
+document.addEventListener('click',e=>{const open=e.target.closest('[data-open]'),source=e.target.closest('[data-source]'),tb=e.target.closest('[data-tab]');if(open){e.preventDefault();openClass(open.dataset.open)}if(source)openSource(source.dataset.source);if(tb)changeTab(tb.dataset.tab)});
+$('#coverageFilter').onchange=e=>{coverageFilter=e.target.value;render()};$('#coverageSort').onchange=e=>{coverageSort=e.target.value;render()};
+$('#hand').onchange=e=>{hand=e.target.value;render()};$('#level').onchange=e=>{level=e.target.value;render()};$('#view').onchange=e=>{view=e.target.value;wirePoses()};$('#reset').onclick=()=>{poses.forEach(p=>p.rot=[0,0]);wirePoses()};$('#closeDetail').onclick=()=>$('#detail').close();$('#detail').addEventListener('close',()=>{activeDetail='';history.replaceState(null,'',location.pathname+location.search)});window.addEventListener('resize',()=>wirePoses());
+function restoreHash(){const hash=location.hash.slice(1);if(hash===activeDetail)return;const initial=new URLSearchParams(hash);if(initial.has('class')&&allClasses[initial.get('class')])openClass(initial.get('class'));else if(initial.has('source')&&D.sources.some(s=>s.id===initial.get('source')))openSource(initial.get('source'));else if(!hash&&$('#detail').open)$('#detail').close()}
+window.addEventListener('hashchange',restoreHash);render();restoreHash();
diff --git a/pose_recluster_distinct_20261014/downloads/coverage/code/web/index.html b/pose_recluster_distinct_20261014/downloads/coverage/code/web/index.html
new file mode 100644
index 0000000000000000000000000000000000000000..485fc151f1175206d27fe439057067f7eef2ebd1
--- /dev/null
+++ b/pose_recluster_distinct_20261014/downloads/coverage/code/web/index.html
@@ -0,0 +1,16 @@
+
+最新抽样 · 局部姿态聚类对比 两组类别 逐类比较 相似与合并 18 行来源表
+WiLoR 外部类 → 我们三源＋合成 外部类筛选 全部 整类未覆盖 未覆盖成员 ≥50% 已覆盖成员 ≥70% 排序 未覆盖比例高 → 低 已覆盖比例高 → 低 类别编号
均指本次抽样库。类别内逐成员统计；未覆盖不等于 GT 可靠，也不等于全量训练集绝对缺失。
判定依据与阈值敏感性
点击外部类的“未覆盖成员 / 原图 / 近邻”查看具体缺口。骨架可拖动旋转；A、B 编号各自独立。
我们：三源＋EgoRenderer Ver.005
类别分布图（仅展示） 同一掌部坐标和投影；图上距离不作为覆盖或合并依据。
右边是外部类；左边是它最相近的我们类。覆盖条来自该类全部成员；点击外部类查看未覆盖成员与各自的精确最近邻。
仍然相近的类 下列类的中心相近，但继续整类合并会把部分差异较大的手指形态连在一起，因此保留并显式标记。
已合并的细类
8 个外部 3D 来源参与本轮聚类。ARCTIC/HOT3D 与我们组重叠；仅 2D 的来源点击直接看 RGB 关键点。
WiLoR 行为官方源数据代理，未声称精确复现其加工训练包。我们组三源依据已核对训练成员的来源 GT；最终 canonical 标签仍未核验。ReInterHand 的 7,600 个 test 手帧未参与聚类。
+
+×
最新抽样 · 局部姿态聚类对比 最新抽样 · 局部姿态聚类对比 两组类别 逐类比较 相似与合并 18 行来源表
-
点击类别查看原图、GT 叠加和最近邻。骨架可拖动旋转；A、B 编号各自独立。
我们：三源＋EgoRenderer Ver.005
类别分布图（仅展示） 同一掌部坐标和投影；图上距离不作为覆盖或合并依据。
右边是外部类；左边是它最相近的我们类。点击可看真实图例；最小样本距离另列，不能把类中心距离当覆盖结论。
WiLoR 外部类 → 我们三源＋合成 外部类筛选 全部 整类未覆盖 未覆盖成员 ≥50% 已覆盖成员 ≥70% 排序 未覆盖比例高 → 低 已覆盖比例高 → 低 类别编号
均指本次抽样库。类别内逐成员统计；未覆盖不等于 GT 可靠，也不等于全量训练集绝对缺失。
判定依据与阈值敏感性
点击外部类的“未覆盖成员 / 原图 / 近邻”查看具体缺口。骨架可拖动旋转；A、B 编号各自独立。
我们：三源＋EgoRenderer Ver.005
类别分布图（仅展示） 同一掌部坐标和投影；图上距离不作为覆盖或合并依据。
右边是外部类；左边是它最相近的我们类。覆盖条来自该类全部成员；点击外部类查看未覆盖成员与各自的精确最近邻。
仍然相近的类 下列类的中心相近，但继续整类合并会把部分差异较大的手指形态连在一起，因此保留并显式标记。
已合并的细类
8 个外部 3D 来源参与本轮聚类。ARCTIC/HOT3D 与我们组重叠；仅 2D 的来源点击直接看 RGB 关键点。
WiLoR 行为官方源数据代理，未声称精确复现其加工训练包。我们组三源依据已核对训练成员的来源 GT；最终 canonical 标签仍未核验。ReInterHand 的 7,600 个 test 手帧未参与聚类。
-
+
×
最新抽样 · 局部姿态聚类对比 两组类别 逐类比较 相似与合并 18 行来源表
+
点击类别查看原图、GT 叠加和最近邻。骨架可拖动旋转；A、B 编号各自独立。
我们：三源＋EgoRenderer Ver.005
类别分布图（仅展示） 同一掌部坐标和投影；图上距离不作为覆盖或合并依据。
右边是外部类；左边是它最相近的我们类。点击可看真实图例；最小样本距离另列，不能把类中心距离当覆盖结论。
仍然相近的类 下列类的中心相近，但继续整类合并会把部分差异较大的手指形态连在一起，因此保留并显式标记。
已合并的细类
8 个外部 3D 来源参与本轮聚类。ARCTIC/HOT3D 与我们组重叠；仅 2D 的来源点击直接看 RGB 关键点。
WiLoR 行为官方源数据代理，未声称精确复现其加工训练包。我们组三源依据已核对训练成员的来源 GT；最终 canonical 标签仍未核验。ReInterHand 的 7,600 个 test 手帧未参与聚类。
+
+×
