from contextlib import closing
from copy import deepcopy

import pytest

from pallet.mask_gallery_review import ReviewError
from pallet.mask_model_training_data import sha
from pallet.pallet_review_triage import select_queue,summary,decide,decisions
from test_pallet_description import editor,item,body


def fixture(n=100):
    classes={'1':dict(class_id=1,fine_index=0,train_enabled=True),'2':dict(class_id=2,fine_index=1,train_enabled=True),'3':dict(class_id=3,fine_index=2,train_enabled=True)}
    props=[dict(target_key=str(i),class_id=1,image_sha256='image'+str(i),review_reasons=['独立分支mask与当前轮廓IoU低于0.85']) for i in range(n)]
    records=[dict(target_key=str(i),class_id=1 if i%2 else 2,evaluation_branch_split='train',component=str(i%10),geometry_source='older') for i in range(n)]
    pred=[dict(target_key=str(i),predicted_fine_index=0,fine_probabilities=[.8,.1,.1],iou=[.3,.6,.75,.82,.95][i%5]) for i in range(n)]
    return props,records,pred,classes,{}


def test_only_conflicts_or_covered_high_score_class_are_priority():
    p,r,m,c,v=fixture();p[0]['review_reasons']=['结构选项／文字／类别定义冲突'];m[1].update(predicted_fine_index=1,fine_probabilities=[.001,.999,0])
    m[2].update(predicted_fine_index=2,fine_probabilities=[.001,0,.999]);m[3].update(predicted_fine_index=1,fine_probabilities=[.02,.98,0])
    q=select_queue(p,r,m,c,v)
    assert {k for k,e in q['entries'].items() if e['tier']=='priority'}=={'0','1'}
    assert not q['entries']['2']['evidence']['training_covered'] and q['actual_audit_count']==50


def test_budget_unique_images_no_completed_resampling_and_determinism():
    args=fixture();p,r,m,c,v=args;v['1']=dict(action='correct',revision=3);p[2]['image_sha256']=p[3]['image_sha256'];q=select_queue(*args)
    audit=[e for e in q['entries'].values() if e['tier']=='audit']
    assert len(audit)==50 and len({e['image_sha256'] for e in audit})==50 and q['entries']['1']['tier']=='diagnostic'
    assert q['audit_strata']=={'geometry':25,'general':25}
    assert select_queue(*args)==q
    before=deepcopy(q);counts=summary(q,{e['target_key']:dict(action='correct') for e in audit})
    assert counts['required_pending']==0 and q==before and len(audit)==50


def test_priority_completed_kept_without_marking_diagnostic_complete():
    p,r,m,c,v=fixture();p[0]['review_reasons']=['结构选项／文字／类别定义冲突'];v['0']=dict(action='correct',revision=2)
    q=select_queue(p,r,m,c,v);counts=summary(q,v)
    assert q['entries']['0']['tier']=='priority' and counts['priority_pending']==0 and counts['required_pending']==50
    assert counts['completed']==1 and counts['unfinished']==99


@pytest.mark.parametrize('reason',['独立分支VLM完整JSON格式失败','已有观察文字但VLM未给出有效观察','独立分支VLM与现有结构字段不一致','逐细类语义抽检'])
def test_model_failures_and_old_perclass_samples_are_not_priority(reason):
    p,r,m,c,v=fixture(10)
    for row in p:row['review_reasons']=[reason]
    q=select_queue(p,r,m,c,v,budget=0)
    assert all(e['tier']=='diagnostic' for e in q['entries'].values())


def test_no_permanent_suppression_of_repeated_highconfidence():
    p,r,m,c,v=fixture();r[0]['geometry_source']='latest_saved_human_correct_round5';m[0].update(predicted_fine_index=1,fine_probabilities=[.001,.999,0])
    q=select_queue(p,r,m,c,v);assert q['entries']['0']['tier']=='priority' and q['entries']['0']['recently_reviewed']


def test_small_pool_never_fills_with_duplicates_or_completed():
    args=fixture(5);args[-1]['0']=dict(action='correct',revision=1);q=select_queue(*args)
    assert q['actual_audit_count']==4


def decision_body(ed,p):return dict(binding=ed.binding,source_key=p['target_key'],source_binding=sha(p),expected_revision=1,request_id='a'*32,policy_sha256='policy',decision='model_error',note='明确模型错误',reviewer='复核人1')


def test_append_decision_never_changes_annotation_or_completion(editor):
    ed,p,g=editor;old=ed.save(p['target_key'],body(ed,p));data=decision_body(ed,p);result=decide(ed,data,'policy')
    assert ed.store.get(p['target_key'])==old and len(ed.store.history(p['target_key']))==1
    assert not result['decision']['annotation_completion_performed'] and decisions(ed)[p['target_key']]==result['decision']
    assert decide(ed,data,'policy')==dict(result,replayed=True)


def test_read_decisions_creates_no_table(editor):
    ed,p,g=editor;assert decisions(ed)=={}
    with closing(ed.store.connect()) as db:assert not db.execute("SELECT name FROM sqlite_master WHERE name='triage_decisions'").fetchone()


@pytest.mark.parametrize('change',[{'expected_revision':0},{'expected_revision':2},{'policy_sha256':'wrong'},{'source_binding':'wrong'},{'decision':'fake_complete'},{'note':'x'*1001}])
def test_decision_invalid_or_stale_blocked(editor,change):
    ed,p,g=editor;old=ed.save(p['target_key'],body(ed,p));data=dict(decision_body(ed,p),**change)
    with pytest.raises(ReviewError):decide(ed,data,'policy')
    assert decisions(ed)=={} and ed.store.get(p['target_key'])==old


def test_decision_id_conflict_and_source_rollback(editor):
    ed,p,g=editor;ed.save(p['target_key'],body(ed,p));data=decision_body(ed,p);decide(ed,data,'policy')
    with pytest.raises(ReviewError):decide(ed,dict(data,decision='label_confirmed'),'policy')
    answers=iter([True,False]);g.current=lambda:next(answers)
    with pytest.raises(ReviewError):decide(ed,dict(data,request_id='b'*32),'policy')
    assert len(decisions(ed))==1
