niceeval 0.5.3 → 0.6.0

package/src/report/aggregate.ts

@@ -4,16 +4,11 @@
 // 和重试策略纠缠在一起(eval A=[1]、eval B=[0,0,0] 平铺 = 0.25,两级宏平均 = 0.5)。
 // 自定义维度把同一道题的 attempt 分进不同组时,第一级折叠发生在各组内部。
 
-import {
-  dedupeAttempts,
-  type AttemptHandle,
-  type Selection,
-  type SelectionWarning,
-  type Snapshot,
-} from "../results/index.ts";
+import { dedupeAttempts } from "../results/select.ts";
+import type { AttemptHandle, Selection, SelectionWarning, Snapshot } from "../results/types.ts";
+import { encodeAttemptLocator, type AttemptLocator } from "../results/locator.ts";
 import type {
   Aggregator,
-  AttemptRef,
   DimensionInput,
   Metric,
   MetricCell,
@@ -21,6 +16,7 @@ import type {
   FlagRef,
 } from "./types.ts";
 import { formatMetricValue } from "./format.ts";
+import { evalPrefixPredicate } from "../shared/aggregate.ts";
 
 // 复合键分隔符:NUL 不会出现在 eval id / experimentId / ISO 时间里,拼接键不会串味
 const KEY_SEP = "\u0000";
@@ -28,7 +24,7 @@ const KEY_SEP = "\u0000";
 /** flag 未声明时的组名:不猜,如实归一组。 */
 export const FLAG_UNSET = "(unset)";
 
-/** 计算函数的第一参:选集(warnings 随行)或手工挑的快照数组(没有挑选过程,自然无警告)。 */
+/** 计算函数的第一参:Selection(warnings 随行)或手工挑的快照数组(没有挑选过程,自然无警告)。 */
 export type SnapshotsInput = Selection | Snapshot[];
 
 export function resolveInput(input: SnapshotsInput): {
@@ -58,6 +54,24 @@ export function snapshotKeyOf(snapshot: Snapshot): string {
   return `${snapshot.experimentId} @ ${snapshot.startedAt}`;
 }
 
+/**
+ * 一条 Item 的 AttemptLocator:真实读取路径(openResults() 产出的 handle)恒有
+ * `attempt.locator`;手工构造的测试 fixture 若省略它,按当前身份元组兜底算一份
+ * (与 `loadAttemptEvidence`、`open.ts` 的回填同一口径)——只服务这类场景,不改变真实
+ * 读取路径的行为。
+ */
+export function locatorOf(item: Item): AttemptLocator {
+  return (
+    item.attempt.locator ??
+    encodeAttemptLocator({
+      experimentId: experimentIdOf(item),
+      snapshotStartedAt: item.snapshot.startedAt,
+      evalId: evalIdOf(item),
+      attempt: item.attempt.result.attempt,
+    })
+  );
+}
+
 /**
  * 展平 + 聚合前去重(niceeval/results 的 dedupeAttempts,身份键
  * (experimentId, evalId, attempt, startedAt))。missing-startedAt 的警告不透出:
@@ -77,15 +91,7 @@ export function collectItems(snapshots: Snapshot[]): Item[] {
   return attempts.map((attempt) => ({ attempt, snapshot: snapshotByAttempt.get(attempt)! }));
 }
 
-/**
- * eval id 前缀过滤,同 CLI 位置参数的分段语义(src/runner/discover.ts):
- * "algebra" 匹配自身与 "algebra/..." 子级,不误配 "algebra2";允许 "algebra/" 尾斜杠写法,等价。
- */
-export function evalPrefixPredicate(evals?: string | string[]): (id: string) => boolean {
-  if (evals === undefined) return () => true;
-  const prefixes = (Array.isArray(evals) ? evals : [evals]).map((p) => p.replace(/\/+$/, ""));
-  return (id) => prefixes.some((prefix) => id === prefix || id.startsWith(prefix + "/"));
-}
+export { evalPrefixPredicate };
 
 export function filterItems(items: Item[], evals?: string | string[]): Item[] {
   if (evals === undefined) return items;
@@ -203,13 +209,13 @@ export function displayValue(metric: Metric, value: number | null): string {
 export async function computeCell(metric: Metric, items: Item[]): Promise<MetricCell> {
   // 第一级桶:同一 (eval × 快照) 的 attempt 折成一个题级值
   const buckets = new Map<string, number[]>();
-  const refs: AttemptRef[] = [];
+  const refs: AttemptLocator[] = [];
   let samples = 0;
   for (const item of items) {
     const value = await evaluateMetric(metric, item.attempt);
     if (value === null) continue;
     samples += 1;
-    refs.push(item.attempt.ref); // 证据引用由句柄直供,不反查下标
+    refs.push(locatorOf(item)); // 证据引用由句柄直供,不反查下标
     const bucketKey = `${evalIdOf(item)}${KEY_SEP}${snapshotKeyOf(item.snapshot)}`;
     const bucket = buckets.get(bucketKey);
     if (bucket) bucket.push(value);