DEV-MERGE-LEGACY-BACKEND-2026/internal/academic/dualrun/attribution.go

package dualrun

import (
	"fmt"
	"math"
)

// severity 把归因折算成"严重度"，用于把一条 case 的多行结论合成一条 case 结论。
//
// 取最大值（越严重越优先），保证"有一条真缺陷"不会被"其余都 MATCH"淹没。
func severity(a Attribution) int {
	switch a {
	case AttrEnvFailure:
		return 6
	case AttrDefect:
		return 5
	case AttrOracleDegraded:
		return 4
	case AttrPathDivergence:
		return 3
	case AttrRoundingConvention:
		return 2
	case AttrPrecisionConvergence:
		return 1
	default:
		return 0
	}
}

// worse 返回两者中更严重的归因。
func worse(a, b Attribution) Attribution {
	if severity(b) > severity(a) {
		return b
	}
	return a
}

// AttrInput 是归因所需的一切（避免把 refs 传进纯函数）。
type AttrInput struct {
	Case Case
	// Oracle oracle 侧观测。
	Oracle *OracleResult
	// Target target 侧纯函数结果。
	Target *TargetResult
	// PredictedDetailID / PredictedTier 是 runner 复刻的兜底判定（用于校验 SP 实返是否与理解一致）。
	PredictedDetailID int64
	PredictedTier     string
}

// Attribute 执行**三步分层归因**（顺序不可换），返回逐行差异与 case 结论。
//
// 第 1 步 income：legacy `round(Σcost,3)` vs 迁移后 `Σ ROUND(cost,2)` —— 不等属精度收敛，继续往下看。
// 第 2 步 规则：oracle 实返的 detail_id 与 runner 复刻判定比对 —— 不等是真缺陷（规则选错/层级错）。
// 第 3 步 teacher_reward：不等时用 `Δreward ≈ Δincome × pct/100` 判断是基数不同还是算法不同。
func Attribute(in AttrInput) ([]DiffRow, Attribution, *Step1Income) {
	obs, tgt, c := in.Oracle, in.Target, in.Case

	// ---- 环境层面：拿不到观测就不判定 ----
	if obs == nil || tgt == nil {
		return []DiffRow{{
			CaseID: c.ID, Table: "-", PK: c.RecordID, Column: "-",
			Legacy: "-", Target: "-", Delta: "-", Attribution: AttrEnvFailure,
		}}, AttrEnvFailure, nil
	}
	if !obs.RollbackRestored {
		return []DiffRow{{
			CaseID: c.ID, Table: "t_courses_record", PK: c.RecordID, Column: "(rollback)",
			Legacy: "restored", Target: "NOT restored", Delta: "-", Attribution: AttrEnvFailure,
		}}, AttrEnvFailure, nil
	}

	// ---- 第 1 步：income ----
	legacyIncome := obs.Inputs.IncomeSumCost3
	targetIncome := tgt.Income
	s1 := &Step1Income{
		Legacy: legacyIncome,
		Target: targetIncome,
		Delta:  math.Round((targetIncome-legacyIncome)*1000000) / 1000000,
	}
	s1.EqualCents = centsOf(legacyIncome) == centsOf(targetIncome)
	s1.Explanation = fmt.Sprintf("legacy round(Σcost,3)=%.3f vs 迁移后 ΣROUND(cost,2)=%.3f", legacyIncome, targetIncome)

	// ---- oracle 退化：不可比，直接标记（R14）----
	if obs.Quote == nil {
		return []DiffRow{{
			CaseID: c.ID, Table: "t_course_payment_detail", PK: c.RecordID, Column: "detail_id",
			Legacy: "未命中（退化分支）", Target: fmt.Sprintf("%.2f (无方案→0)", tgt.Reward),
			Delta: "-", Attribution: AttrOracleDegraded,
		}}, AttrOracleDegraded, s1
	}

	// ---- 第 2 步：规则命中是否与复刻判定一致 ----
	var rows []DiffRow
	attr := AttrMatch
	actualDetail := int64(obs.Quote.DetailID)
	if in.PredictedDetailID != 0 && actualDetail != in.PredictedDetailID {
		rows = append(rows, DiffRow{
			CaseID: c.ID, Table: "t_course_payment_detail", PK: c.RecordID, Column: "detail_id",
			Legacy: fmt.Sprintf("%d", actualDetail), Target: fmt.Sprintf("%d", in.PredictedDetailID),
			Delta: "-", Attribution: AttrDefect,
		})
		attr = worse(attr, AttrDefect)
	}
	if c.ExpectTier != "" && in.PredictedTier != "" && c.ExpectTier != in.PredictedTier {
		rows = append(rows, DiffRow{
			CaseID: c.ID, Table: "path", PK: c.RecordID, Column: "tier",
			Legacy: in.PredictedTier, Target: c.ExpectTier, Delta: "-", Attribution: AttrPathDivergence,
		})
		attr = worse(attr, AttrPathDivergence)
	}

	// ---- 第 3 步：teacher_reward ----
	legacyReward := 0.0
	if obs.After != nil {
		legacyReward = obs.After.TeacherReward
	}
	rewardRow := DiffRow{
		CaseID: c.ID, Table: "t_courses_record", PK: c.RecordID, Column: "teacher_reward",
		Legacy: fmt.Sprintf("%.3f", legacyReward),
		Target: fmt.Sprintf("%.2f", tgt.Reward),
		Delta:  fmt.Sprintf("%+.3f", tgt.Reward-legacyReward),
	}

	deltaIncomeCents := centsOf(targetIncome) - centsOf(legacyIncome)
	if centsOf(legacyReward) == centsOf(tgt.Reward) {
		rewardRow.Attribution = AttrMatch
		rows = append(rows, rewardRow)
	} else {
		// 判据顺序（不可换）：
		//   ① Δincome ≠ 0 且金额成比例  ⇒ PRECISION-CONVERGENCE（基数变了）
		//   ② Δincome = 0              ⇒ DEFECT（基数没变而金额变了 = 算法不同）
		//   ③ 其余（不成比例）          ⇒ DEFECT
		expectedDelta := s1.Delta * obs.Quote.Percentage / 100
		actualDelta := tgt.Reward - legacyReward
		tolerance := math.Max(0.015, math.Abs(expectedDelta)+0.015) // 1.5 分容差
		proportional := math.Abs(actualDelta-expectedDelta) <= tolerance
		switch {
		case deltaIncomeCents != 0 && proportional:
			rewardRow.Attribution = AttrPrecisionConvergence
			rewardRow.Column += fmt.Sprintf("(Δincome=%+d分, Δ预期 %+.4f, 实测 %+.4f)", deltaIncomeCents, expectedDelta, actualDelta)
			attr = worse(attr, AttrPrecisionConvergence)
		case deltaIncomeCents == 0:
			rewardRow.Attribution = AttrDefect
			rewardRow.Column += fmt.Sprintf("(Δincome=0 而 Δreward=%+.4f ⇒ 基数相同、算法不同)", actualDelta)
			attr = worse(attr, AttrDefect)
		default:
			rewardRow.Attribution = AttrDefect
			rewardRow.Column += fmt.Sprintf("(Δincome=%+d分, Δ预期 %+.4f, 实测 %+.4f, 不成比例)", deltaIncomeCents, expectedDelta, actualDelta)
			attr = worse(attr, AttrDefect)
		}
		rows = append(rows, rewardRow)
	}

	// ---- 逐行分摊（R13：不参与通过判定，但必须逐行记录）----
	rows = append(rows, attributePerRow(c, obs, tgt, attr)...)
	return rows, attr, s1
}

// attributePerRow 逐行比对；差异按 R13 归因（PHP 每行 round(,3) 不守恒 vs Go 分位分配守恒）。
func attributePerRow(c Case, obs *OracleResult, tgt *TargetResult, caseAttr Attribution) []DiffRow {
	legacy := map[int64]float64{}
	for _, s := range obs.StudentsAfter {
		if s.Type == 1 || s.Type == 2 {
			legacy[s.StudentID] = s.TeacherReward
		}
	}
	var rows []DiffRow
	legacySum := 0.0
	for _, v := range legacy {
		legacySum += v
	}
	targetSum := 0.0
	for _, v := range tgt.PerRow {
		targetSum += v
	}
	totalLegacy := 0.0
	if obs.After != nil {
		totalLegacy = obs.After.TeacherReward
	}

	for i, sid := range tgt.RowStudents {
		lv, ok := legacy[sid]
		if !ok {
			continue
		}
		tv := tgt.PerRow[i]
		if centsOf(lv) == centsOf(tv) {
			continue // 逐行相等就不刷行，避免报告噪声
		}
		rowAttr := AttrDefect
		deltaIncomeCents := centsOf(tgt.Income) - centsOf(obs.Inputs.IncomeSumCost3)
		switch {
		case deltaIncomeCents == 0:
			// Δincome = 0 而逐行不等 ⇒ 只可能是分摊方式不同（R13）
			rowAttr = AttrRoundingConvention
		case math.Abs(legacySum-totalLegacy) > 0.0005 && math.Abs(targetSum-totalLegacy) < 0.005:
			// PHP 逐行和不等于 record 总额（不守恒）、Go 精确守恒 ⇒ 舍入口径 + 基数变化
			rowAttr = AttrRoundingConvention
		case caseAttr == AttrPrecisionConvergence:
			rowAttr = AttrPrecisionConvergence
		}
		rows = append(rows, DiffRow{
			CaseID: c.ID, Table: "t_courses_record_student", PK: fmt.Sprintf("%s/%d", c.RecordID, sid),
			Column: "teacher_reward",
			Legacy: fmt.Sprintf("%.3f", lv), Target: fmt.Sprintf("%.2f", tv),
			Delta: fmt.Sprintf("%+.3f", tv-lv), Attribution: rowAttr,
		})
	}
	return rows
}

// CaseAttribution 把一行行结论合成 case 结论（取最严重）。
func CaseAttribution(rows []DiffRow, fallback Attribution) Attribution {
	out := fallback
	for _, r := range rows {
		out = worse(out, r.Attribution)
	}
	return out
}