"""Exact numerical gates and measured performance report for local probes."""
from pathlib import Path
import hashlib, json, statistics, subprocess

ROOT=Path(__file__).resolve().parents[2]
OUT=ROOT/'test/local-probe-20260917'
def read(path):return json.loads((OUT/path).read_text(encoding='utf-8'))
def table(headers,rows):return '\n| '+' | '.join(headers)+' |\n| '+' | '.join(['---']*len(headers))+' |\n'+'\n'.join('| '+' | '.join(map(str,row))+' |' for row in rows)+'\n'

def compare_matrices(left,right,n):
    size=8*(1+n+n*n);checked=0
    with left.open('rb') as a,right.open('rb') as b:
        while True:
            x,y=a.read(size),b.read(size)
            if not x and not y:break
            assert len(x)==len(y)==size,('matrix record size',checked)
            assert x[:8*(1+n)]==y[:8*(1+n)],('Jacobian input t/y differs',checked)
            assert x[8*(1+n):]==y[8*(1+n):],('Jacobian matrix bits differ',checked)
            checked+=1
    return dict(matrices=checked,entries=checked*n*n,differentEntries=0,inputTimesAndStatesIdentical=True)

def main():
    plan=read('plan.json');reference=read('reference-audit/measurement.json')
    cases=['group26-final-audit','small-final-audit','all-audit']
    exact=['statesSha256','outputsSha256','eventsSha256','finalState','final','propertyWarnings',
           'acceptedSteps','rejectedSteps','stateTransitions','solverStarts','nfev','njev','nlu']
    checks=[]
    for path in sorted(OUT.glob('*/measurement.json')):
        obj=json.loads(path.read_text(encoding='utf-8'))
        differences=[k for k in exact if obj[k]!=reference[k]]
        for k in ('newtonIterations','newtonConvergenceFailures'):
            if obj['diagnostic'][k]!=reference['diagnostic'][k]:differences.append(k)
        checks.append(dict(run=path.parent.name,differences=differences))
    assert not any(x['differences'] for x in checks),checks
    matrices={name:compare_matrices(OUT/'reference-audit/jacobians.bin',OUT/name/'jacobians.bin',len(plan['stateKeys'])) for name in cases}
    allrun=read('all-audit/measurement.json');diag=allrun['diagnostic'];groups=[]
    for theory,actual in zip(plan['groups'],diag['groups']):
        assert all(a+b==896 for a,b in zip(actual['executed'],actual['skipped']))
        assert not any(actual['skipped'][i] for i in theory['affectedOperations'])
        groups.append({**theory,**actual,'executedTotal':sum(actual['executed']),'skippedTotal':sum(actual['skipped'])})
    metrics=['solveCpuSeconds','solveSeconds','processSeconds','jacobianSeconds']
    measurements=[];medians={}
    for name in ('reference','all'):
        runs=[read(f'{name}-run-{i}/measurement.json') for i in range(5)]
        for i,r in enumerate(runs):
            measurements.append(dict(mode=name,run=i,**{k:r[k] for k in metrics[:-1]},jacobianSeconds=r['diagnostic']['jacobianSeconds']))
        medians[name]={k:statistics.median(r['diagnostic'][k] if k=='jacobianSeconds' else r[k] for r in runs) for k in metrics}
    reductions={k:1-medians['all'][k]/medians['reference'][k] for k in metrics}
    source_meta=read('audit/build-metadata.json')
    source_unchanged=all(hashlib.sha256((ROOT/'native'/key).read_text(encoding='utf-8').encode()).hexdigest()==value for key,value in source_meta['sourceHashes'].items())
    assert source_unchanged
    # The original evaluator, ordinary residual, temperature checker and linear
    # solver are not patched. The added evaluator is reachable only from lp_eval.
    original=(OUT/'original-model.c').read_text(encoding='utf-8')
    for mode in ('audit','worker'):
        emitted=(OUT/mode/'model.c').read_text(encoding='utf-8')
        assert original in emitted,'original model source changed'
    old=ROOT/'test/probe-state-residual-20260916/baseline-run-1'
    prior={}
    for name in ('states','outputs'):
        if (old/f'{name}.bin').exists():
            with (old/f'{name}.bin').open('rb') as f:prior[name]=hashlib.file_digest(f,'sha256').hexdigest()==reference[name+'Sha256']
    assert all(prior.values())
    result=dict(numericalChecks=checks,matrixComparisons=matrices,
        shadow=dict(comparisons=diag['auditComparisons'],differentComparisons=diag['auditDifferences'],
                    comparedDerivativeAndOutputValues=diag['auditComparisons']*(len(plan['stateKeys'])+len(reference['final']))),
        originalNativeSourcesUnchanged=source_unchanged,originalGeneratedModelVerbatim=True,priorUninstrumentedOutputsEqual=prior,
        medians=medians,reductions=reductions,measurements=measurements,groups=groups,
        workspaceBytes=diag['workspaceBytes'],copiedBytes=diag['contextCopiedBytes'],comparedBytes=diag['contextComparedBytes'])
    (OUT/'comparison.json').write_text(json.dumps(result,ensure_ascii=False,indent=2)+'\n',encoding='utf-8')
    report=[]
    report.append('**局部 Jacobian probe 实验结果（2026-09-17）**\n')
    report.append('本次增加独立实验构建工具，未接入前端或生产默认路径。生产 ordinary residual、物性算法、accepted-step check、线性求解器及原始生成模型源码保持不变；实验 worker 只在 Jacobian callback 中选择额外生成的局部求值函数。对象为 `tests/data/test-mql-8-corrected.json`，原提交 `1aac220`，CVODE 7.4.0，0–10 s，BDF，rtol=1e-8，沿用原 atol、步长和线性求解设置。\n')
    report.append('**实施范围与保护**\n')
    report.append('从现有 StateDependencies 和 EvaluationSchedule 保守计算各 group 的状态依赖闭包，未解析用户 C 或手写组件白名单。未知依赖按全部状态处理；当前实验拒绝包含循环代数块的模型。每组生成连续的不受扰动调度区间。扰动时间和组外状态必须与本次 Jacobian baseline 逐位一致，跨 Jacobian 基准失效。\n')
    report.append('原型只裁剪 flow/stream/linear/alias 调度段。气体状态准备、后续节点能量、端口传播、机械方程、各管路诊断输出和有限值检查继续执行原代码。纯代数区间依据闭包复制基准的所有区间输出；包含上下文调用的区间，还要求入口的完整 live property states、count/capacity、observer/Jacobian 指针及全部 pipe cache 与 baseline 逐位一致，再恢复基准出口上下文和所有区间输出。任何不一致均执行原计算，无近似匹配、无放宽容差。\n')
    report.append('保留 property context 中的 h/rho/黏度/等熵派生字段及种子顺序，避免只按显式 p,T 或 p,h 误判等价。已有 Jacobian 标量表在 baseline 后保持只读，跳过调用只减少其评估/复用统计，不改变表内数值。温度观察路径不走局部函数。\n')
    report.append('初版第26组通过数值验证，但无条件保存所有快照导致约5.11 GB复制。最终版按启用组选择快照，并在纯代数操作之间共享上下文检查点：321个调度边界只需57个上下文检查点。只共享确定不含上下文调用的检查点；未放宽任何数值或上下文比较条件。\n')
    report.append('**先验证数值等价**\n')
    report.append(table(['验证阶段','完整矩阵数','逐位比较矩阵元素数','矩阵差异','每次Jacobian求值的dy/w对照'],[(name,matrices[name]['matrices'],matrices[name]['entries'],0,'25,088次，差异0') for name in cases]))
    report.append('每份矩阵记录同时比较 t、132维输入 y 和132×132矩阵的原始双精度位，包含稀疏结构之外的零元素。另用独立 Jacobian memo workspace 在同一(t,y)执行原实现，比较每次 baseline/probe 的全部132个dy及1784个w；任意差异立即终止。该影子计算只在 audit worker 启用，不计入求解器nfev；性能 worker 完全关闭影子计算、矩阵落盘及内核入口计数。\n')
    report.append(table(['项目','原实现','局部全部group','结果'],[
        ('states / outputs 原始二进制','完整文件','完整文件','逐字节一致'),('warning','全部告警字段','全部告警字段','一致'),
        ('event','事件时刻及132维事件后状态','同左','逐位一致'),
        *[(k,reference[k],allrun[k],'一致') for k in ('acceptedSteps','rejectedSteps','stateTransitions','solverStarts','nfev','njev','nlu')],
        ('Newton iterations',reference['diagnostic']['newtonIterations'],diag['newtonIterations'],'一致'),
        ('Newton convergence failures',reference['diagnostic']['newtonConvergenceFailures'],diag['newtonConvergenceFailures'],'一致')]))
    report.append('所有已保存运行的完整结果均对照通过；还核对了前一轮真正未插桩 worker 的 states/outputs，两者逐字节相同。图闭包及循环块拒绝的5项单元检查通过；原生 guard 测试验证了同p,T但派生rho改变、pipe cache改变、上下文长度/指针改变、t改变、组外状态改变和跨基准失效均不会误命中，并验证纯代数区间恢复不会覆盖当前物性上下文。未发现需要用容差掩盖的依赖裁剪差异。\n')
    report.append('**实际执行范围**\n')
    total=sum(g['executedTotal']+g['skippedTotal'] for g in groups);skipped=sum(g['skippedTotal'] for g in groups)
    report.append(f'Jacobian次数仍为896，每个Jacobian仍调用1次baseline及27次probe，合计25,088次，其中probe为24,192次。probe 调度操作从 **{total:,}** 次降为 **{total-skipped:,}** 次，实际跳过 **{skipped:,}（{skipped/total:.2%}）**。与此前88.12%的结构无关操作比例相比，完整上下文保护、保留的诊断/后续操作限制了可安全跳过的范围。\n')
    report.append(table(['group','扰动状态数','图上受影响操作数/484','实际执行操作/896次probe','跳过比例','上下文保护回退次数'],[(g['color'],len(g['stateIndices']),len(g['affectedOperations']),g['executedTotal'],f'{g["skippedTotal"]/(484*896):.2%}',g['contextMisses']) for g in groups]))
    report.append('每组的被扰动状态、受影响操作编号、每一项操作的执行和跳过计数均保存在 `comparison.json`；操作编号、名称、输入输出和依赖状态在 `plan.json`。所有图上可能受影响操作的跳过次数均为0。下面统计24,192次probe实际进入内核的次数，不含baseline、影子计算或普通residual。\n')
    kernel_rows=[]
    for i,name in enumerate(diag['kernelNames']):
        a=sum(row[i] for row in reference['diagnostic']['kernelCalls'][1:]);b=sum(row[i] for row in diag['kernelCalls'][1:])
        kernel_rows.append((name,a,b,f'{1-b/a:.2%}' if a else '—'))
    report.append(table(['实际计算','原实现','局部probe','减少'],kernel_rows))
    report.append('gas、PH、density和pipe root的实际调用未变；已有精确缓存本来就已避免大量无关求根。此次节省来自流量、等熵、黏度及属性查找的实际执行。管路诊断仍全部执行，未将其删除或替换为未验证的近似值。\n')
    report.append('**实际耗时**\n')
    report.append('同一不含影子核验的可执行文件，通过 LOCAL_PROBE_MASK=0 和全组mask切换；排除一次预热，5对交替顺序串行测量，每次均完整输出states/outputs。Jacobian为整个callback的QPC墙钟时间；积分CPU为现有整进程CPU计时；完整native仿真时间从启动worker到输出完成并退出，包含DLL加载、初始化、积分、采样、输出重放/编码/落盘及退出。该数字不包含浏览器、API排队或已完成的编译。\n')
    names={'jacobianSeconds':'Jacobian阶段墙钟','solveCpuSeconds':'积分CPU','solveSeconds':'积分墙钟','processSeconds':'完整native仿真墙钟'}
    report.append(table(['中位数','原实现 / s','局部全部group / s','减少'],[(names[k],f'{medians["reference"][k]:.6f}',f'{medians["all"][k]:.6f}',f'{reductions[k]:.2%}') for k in ('jacobianSeconds','solveCpuSeconds','solveSeconds','processSeconds')]))
    report.append(table(['序号/模式','Jacobian / s','积分CPU / s','积分墙钟 / s','完整native / s'],[(f'{r["run"]}/{r["mode"]}',*[f'{r[k]:.6f}' for k in ('jacobianSeconds','solveCpuSeconds','solveSeconds','processSeconds')]) for r in measurements]))
    report.append('全部样本保留，包括原实现第2次较慢的样本，没有按有利结果筛选。当前测量显示收益，但绝对秒数受机器负载和频率影响；不能把前一日7.07s与今日结果直接相减归为优化收益。\n')
    report.append(f'最终局部工作区占 **{diag["workspaceBytes"]:,} 字节**。整轮额外上下文复制 **{diag["contextCopiedBytes"]:,} 字节**、比较请求覆盖 **{diag["contextComparedBytes"]:,} 字节**；memcmp可能提前终止，后者不是硬件实际读取字节数。上述机制的开销已经包含在性能测量中。还包括switch分派、组外状态检查和恢复输出。\n')
    cold=[]
    for name in ('reference','all'):
        path=OUT/f'cold-{name}-pipeline.json'
        if path.exists():cold.append(json.loads(path.read_text()))
    if cold:
        report.append('另外从工程JSON开始执行实验的冷构建完整链（读取/方程生成/实验代码生成/GCC编译/DLL准备/worker运行/输出），实测如下。两种运行使用同一实验构建机制，冷编译时间会波动；这不是前端API链路测量，也不能用单次冷启动数据判断probe收益。\n')
        report.append(table(['冷构建链','准备与构建 / s','完整native / s','完整实验流水线 / s'],[(x['mode'],f'{x["prepareSeconds"]:.6f}',f'{x["workerSeconds"]:.6f}',f'{x["pipelineSeconds"]:.6f}') for x in cold]))
    report.append('**边界与后续方向**\n')
    report.append('此版本可作为验证原型，暂不建议直接接入默认求解路径：仅支持当前无循环调度的生成模型，快照内存随模型规模增长，实验状态为独立worker内的静态运行期对象，不适合直接嵌入同进程并行仿真。较大group的属性上下文经常变化，保守整上下文比较会使本可独立的区间回退，收益小于最初理论无关计算预算。\n')
    report.append('下一步若继续推进，应先研究保留完整读写语义的更细粒度context依赖，再扩展到诊断/其他阶段。不能直接删除上下文保护，也不能把88.12%的无关操作比例解释为可实现的同等加速。当前实验已证明可以在不改变数值轨迹的前提下减少一部分probe计算，但没有消除全部无关执行。\n')
    (OUT/'report.md').write_text('\n'.join(report),encoding='utf-8')
    print(json.dumps(dict(medians=medians,reductions=reductions,matrices=matrices,skippedOperations=skipped,operationReduction=skipped/total,numericalRuns=len(checks)),ensure_ascii=False,indent=2))

if __name__=='__main__':main()