import io from pathlib import Path from types import SimpleNamespace import numpy as np import pytest from scipy.io import wavfile from external_visual import JoinedReader, exclusion from session_overlay import ExternalOverlay, timing from common import sha256 def test_split_archive_reader(tmp_path): paths=[] for i,b in enumerate((b'abc',b'',b'defgh')): p=tmp_path/str(i);p.write_bytes(b);paths.append(p) with JoinedReader(paths) as stream, io.BufferedReader(stream,2) as buffer: assert buffer.read(4)==b'abcd' assert buffer.read()==b'efgh' def test_timestamp_not_relocated_to_fit(): row=dict(question_time=6.13,duration_seconds=20) assert timing(row,[(3,9)],2,12,11.8)==(4.13,6.13) assert timing(row,[(1,4),(7,12)],2,12,11.8) is None assert timing({**row,'duration_seconds':12},[(3,9)],2,12,11.8) is None assert timing(row,[(3,9)],2,12,6.2) is None # room for reply windows def test_original_future_tasks_never_admitted(): assert exclusion(dict(task='FPD'))=='anticipation_or_proactive_task_not_prefix_qa' assert exclusion(dict(task='CRR'))=='anticipation_or_proactive_task_not_prefix_qa' assert exclusion(dict(task='EPM',video='Ego4D/clips/x.mp4'))=='source_access_not_confirmed' def test_overlay_keeps_motion_controls_and_original_events(tmp_path): wave=(np.sin(np.arange(32000)*.05)*2000).astype(np.int16) path=tmp_path/'question.wav';wavfile.write(path,16000,wave) steps=[dict(step_index=i,semantic_window_index=i//20,update_semantic_cache=i%20==0, audio_waveform=np.zeros(9600,np.float32),states=np.zeros(36,np.float32), actions=np.full((30,40),i,np.float32),control_label=1) for i in range(600)] original=dict(source_id='immutable_motion',steps=steps,session_voice_name='female_01_en', interaction_mode='turn-taking',microphone_speech_intervals=[dict(start_seconds=0,end_seconds=2)]) overlay=ExternalOverlay(SimpleNamespace(motion_lead_seconds=1.,motion_tail_seconds=.8,question_fade_seconds=.02)) overlay.by_voice={'female_01_en':[dict(question_time=6.13,duration_seconds=20, asset_id='ovo_test',question_audio_path=str(path),question_audio_sha256=sha256(path), video_path='video.mp4',visual_features_path='features.pt',question='What is visible?',answer='A cup.', data_domain='OVO-Bench',external_id='ovo_test',original_question='What is visible?', original_answer='A cup.',options=['A cup.','A bowl.'],correct_option=0, video_sha256='sha',visual_receipt='review.json')]} result=overlay.compose(original) assert result['multimodal_overlay']['mode']=='qa' assert abs(result['multimodal_overlay']['question_end']-6.13)<1/16000 assert result['microphone_speech_intervals']==original['microphone_speech_intervals'] for before,after in zip(steps,result['steps']): assert after['actions'] is before['actions'] assert after['states'] is before['states'] assert after['control_label']==before['control_label'] assert not before['audio_waveform'].any() assert any(s['audio_waveform'].any() for s in result['steps']) def test_unselected_cleanup_is_exact_and_preserves_other_files(tmp_path): from prepare import discard_unselected row=dict(dataset_index=7,sample_id='hfd_session_00007') for suffix in ('.json','.pt','_microphone.wav','_reference.npz'): (tmp_path/(row['sample_id']+suffix)).write_bytes(b'generated') keep=tmp_path/'keep.txt';keep.write_text('user evidence') discard_unselected(row,tmp_path) assert list(tmp_path.iterdir())==[keep] with pytest.raises(ValueError): discard_unselected(dict(dataset_index=7,sample_id='../keep'),tmp_path) def test_revision_refuses_changed_inputs(tmp_path): from revision import link_checked, same_asset source=tmp_path/'old';source.write_bytes(b'original source') dest=tmp_path/'new' with pytest.raises(ValueError):link_checked(source,dest,'wrong hash') assert not dest.exists() link_checked(source,dest,sha256(source)) assert source.stat().st_ino==dest.stat().st_ino asset=dict(asset_id='q',voice='v',video_sha256='video',visual_features_sha256='feature', question_audio_sha256='pcm',visual_receipt_sha256='review',question_time=5, question='Q',answer='A',original_question='Q',original_answer='A') row=dict(voice='v',overlay=dict(asset_id='q',question='Q',answer='A',external_benchmark=dict( source_video_sha256='video',feature_sha256='feature',question_audio_sha256='pcm', review_receipt_sha256='review',question_time=5,original_question='Q',original_answer='A'))) assert same_asset(row,asset) assert not same_asset(row,{**asset,'answer':'different'}) assert not same_asset(row,{**asset,'question_time':6}) def test_revision_reuses_only_complete_identical_results(tmp_path): import json from common import write_json from revision import reuse_generated old=tmp_path/'old';new=tmp_path/'new';out=tmp_path/'old_outputs';dest=tmp_path/'new_outputs' for folder in (old,new,out):folder.mkdir() rows=[] for sid in ('unchanged','changed'): row=dict(sample_id=sid) for field,suffix in [('source','pt'),('reference','npz'),('microphone','wav')]: name=sid+'_'+field+'.'+suffix (old/name).write_bytes((sid+field).encode());(new/name).write_bytes((sid+field).encode()) row[field+'_file']=name;row[field+'_sha256']=sha256(old/name) rows.append(row) previous=dict(run='same',run_config_sha256='config',rows=rows) write_json(old/'manifest.json',previous) after=json.loads(json.dumps(previous));after['rows'][1]['source_sha256']='different' write_json(new/'manifest.json',after) identity=dict(cohort_sha256=sha256(old/'manifest.json'),run='same',step=1,native_speech=True) write_json(out/'identity.json',identity) for row in rows: sid=row['sample_id'];(out/(sid+'.npz')).write_bytes(b'motion') write_json(out/(sid+'.json'),dict(**identity,answers=[])) reuse_generated(old,new,out,dest) assert (dest/'unchanged.npz').exists() and not (dest/'changed.npz').exists() result=json.loads((dest/'unchanged.json').read_text()) assert result['cohort_sha256']==sha256(new/'manifest.json') assert result['reuse_provenance']['original_cohort_sha256']==sha256(old/'manifest.json') assert json.loads((out/'unchanged.json').read_text())==dict(**identity,answers=[])