fix(routing): 对齐缓存亲和力硬规则与审计

apps/api/internal/store/candidates.go

@@ -500,14 +500,15 @@ func applyRuntimeCandidateCacheAffinity(candidate *RuntimeModelCandidate, option
 		key = options.CacheAffinityKey
 	}
 	affinity := RuntimeCandidateCacheAffinity{
-		Key:               key,
-		RequestCount:      input.RequestCount,
-		InputTokens:       input.InputTokens,
-		CachedInputTokens: input.CachedInputTokens,
-		EMAHitRatio:       boundedRatio(input.EMAHitRatio),
-		LastHitRatio:      boundedRatio(input.LastHitRatio),
-		LastObservedUnix:  input.LastObservedUnix,
-		AdjustedPriority:  float64(candidate.PlatformPriority),
+		Key:                key,
+		RequestCount:       input.RequestCount,
+		InputTokens:        input.InputTokens,
+		CachedInputTokens:  input.CachedInputTokens,
+		EMAHitRatio:        boundedRatio(input.EMAHitRatio),
+		LastHitRatio:       boundedRatio(input.LastHitRatio),
+		LastObservedUnix:   input.LastObservedUnix,
+		AdjustedPriority:   float64(candidate.PlatformPriority),
+		MatchedPrefixDepth: cacheAffinityMatchedPrefixDepth(key, options.CacheAffinityKeys),
 	}
 	minSamples := cacheAffinityMinSamples(options.CacheAffinityPolicy)
 	hasObservedCachedHit := affinity.CachedInputTokens > 0 || affinity.LastHitRatio > 0
@@ -538,6 +539,32 @@ func applyRuntimeCandidateCacheAffinity(candidate *RuntimeModelCandidate, option
 	candidate.CacheAffinity = affinity
 }
 
+func cacheAffinityMatchedPrefixDepth(key string, lookup []string) int {
+	prefix := ""
+	switch {
+	case strings.HasPrefix(key, "prompt_lcp_v2:"):
+		prefix = "prompt_lcp_v2:"
+	case strings.HasPrefix(key, "prompt_lcp:"):
+		prefix = "prompt_lcp:"
+	default:
+		return 0
+	}
+	matchingKeys := make([]string, 0, len(lookup))
+	for _, candidateKey := range lookup {
+		if strings.HasPrefix(candidateKey, prefix) {
+			matchingKeys = append(matchingKeys, candidateKey)
+		}
+	}
+	for index, candidateKey := range matchingKeys {
+		if candidateKey == key {
+			return len(matchingKeys) - index + cacheAffinityMinimumPrefixDepth - 1
+		}
+	}
+	return 0
+}
+
+const cacheAffinityMinimumPrefixDepth = 2
+
 func cacheAffinityPolicyEnabled(policy map[string]any, modelType string) bool {
 	if enabled, ok := policy["enabled"].(bool); ok && !enabled {
 		return false
@@ -712,6 +739,9 @@ func sortRuntimeModelCandidates(items []RuntimeModelCandidate) {
 		if aFull != bFull {
 			return !aFull
 		}
+		if items[i].PlatformPriority != items[j].PlatformPriority {
+			return items[i].PlatformPriority < items[j].PlatformPriority
+		}
 		if items[i].CacheAffinity.Applied != items[j].CacheAffinity.Applied {
 			return items[i].CacheAffinity.Applied
 		}
@@ -746,6 +776,32 @@ func sortRuntimeModelCandidates(items []RuntimeModelCandidate) {
 		}
 		return false
 	})
+	appliedCount := 0
+	for index := range items {
+		if items[index].CacheAffinity.Applied {
+			appliedCount++
+		}
+	}
+	for index := range items {
+		items[index].CacheAffinity.CandidateCount = appliedCount
+	}
+	if len(items) == 0 || items[0].CacheAffinity.Applied {
+		return
+	}
+	for index := 1; index < len(items); index++ {
+		if !items[index].CacheAffinity.Applied {
+			continue
+		}
+		switch {
+		case runtimeCandidateFull(items[index]) && !runtimeCandidateFull(items[0]):
+			items[0].CacheAffinity.OverrideReason = "capacity_tier_unavailable"
+		case items[index].PlatformPriority != items[0].PlatformPriority:
+			items[0].CacheAffinity.OverrideReason = "different_effective_priority_tier"
+		}
+		if items[0].CacheAffinity.OverrideReason != "" {
+			break
+		}
+	}
 }
 
 func runtimeCandidateFull(candidate RuntimeModelCandidate) bool {