niceeval 0.15.1-canary.76 → 0.16.1

package/dist/runner/reuse-diagnostics.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"reuse-diagnostics.js","sourceRoot":"","sources":["reuse-diagnostics.ts"],"names":[],"mappings":";;AAiCA,4DA+CC;AAED,8DAMC;AAjFD,iDAAiD;AACjD,MAAM,WAAW,GAAG,CAAC,CAAC;AAatB,SAAS,OAAO,CAAC,MAAkB;IAC/B,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS,CAAC;AACvE,CAAC;AACD,yEAAyE;AACzE,SAAS,OAAO,CAAC,MAAkB;IAC/B,OAAO,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,UAAU,CAAC;AAC5G,CAAC;AACD;;;;GAIG;AACH,SAAgB,wBAAwB,CAAC,OAA8B;IACnE,MAAM,UAAU,GAAG,IAAI,GAAG,EAItB,CAAC;IACL,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC3B,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;QAC/B,IAAI,CAAC,OAAO,EAAE,MAAM,IAAI,OAAO,CAAC,YAAY,KAAK,SAAS,IAAI,OAAO,CAAC,YAAY,KAAK,SAAS;YAC5F,SAAS;QACb,MAAM,GAAG,GAAG,GAAG,MAAM,CAAC,YAAY,IAAI,EAAE,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QACnE,IAAI,KAAK,GAAG,UAAU,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAChC,IAAI,CAAC,KAAK,EAAE,CAAC;YACT,KAAK,GAAG,EAAE,GAAG,CAAC,MAAM,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,MAAM,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,YAAY,EAAE,OAAO,CAAC,YAAY,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;YAClJ,UAAU,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QAC/B,CAAC;QACD,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAChC,CAAC;IACD,MAAM,OAAO,GAA+B,EAAE,CAAC;IAC/C,KAAK,MAAM,KAAK,IAAI,UAAU,CAAC,MAAM,EAAE,EAAE,CAAC;QACtC,MAAM,KAAK,GAAG,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAQ,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC;QACxE,kDAAkD;QAClD,IAAI,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC;YACxB,SAAS;QACb,MAAM,aAAa,GAAG,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAQ,CAAC,YAAa,IAAI,CAAC,IAAI,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAChG,MAAM,OAAO,GAAG,IAAI,GAAG,EAA4B,CAAC;QACpD,KAAK,MAAM,OAAO,IAAI,aAAa,EAAE,CAAC;YAClC,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;YAC/B,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;YAC1C,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,OAAQ,CAAC,YAAa,CAAC,CAAC;YAC9C,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;QACjC,CAAC;QACD,KAAK,MAAM,CAAC,KAAK,EAAE,QAAQ,CAAC,IAAI,OAAO,EAAE,CAAC;YACtC,IAAI,QAAQ,CAAC,MAAM,GAAG,WAAW;gBAC7B,SAAS;YACb,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YACnD,OAAO,CAAC,IAAI,CAAC;gBACT,GAAG,CAAC,KAAK,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,KAAK,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACjF,YAAY,EAAE,KAAK,CAAC,YAAY;gBAChC,KAAK;gBACL,WAAW,EAAE,MAAM,CAAC,CAAC,CAAE;gBACvB,SAAS,EAAE,MAAM,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAE;gBACrC,KAAK,EAAE,MAAM,CAAC,MAAM;aACvB,CAAC,CAAC;QACP,CAAC;IACL,CAAC;IACD,OAAO,OAAO,CAAC;AACnB,CAAC;AACD,kCAAkC;AAClC,SAAgB,yBAAyB,CAAC,MAAgC;IACtE,MAAM,KAAK,GAAG,MAAM,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,mBAAmB,MAAM,CAAC,YAAY,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;IACjG,OAAO,CAAC,iCAAiC,MAAM,CAAC,YAAY,GAAG,KAAK,0CAA0C;QAC1G,GAAG,MAAM,CAAC,KAAK,4BAA4B,MAAM,CAAC,WAAW,IAAI,MAAM,CAAC,SAAS,oBAAoB,MAAM,CAAC,KAAK,IAAI;QACrH,6GAA6G;QAC7G,4GAA4G,CAAC,CAAC;AACtH,CAAC","sourcesContent":["// 复用污染的收尾诊断(契约见 docs/feature/sandbox/reuse.md「复用污染的可观察性」)。\n//\n// 「setup 幂等、不依赖 workdir 外残留」是作者义务,违约的症状(下游 Eval 莫名失败)不指向复用。\n// Run 收尾按 Sandbox 实例与承接序号聚合:首承接正常、而同一实例序号 ≥ 2 的 Attempt 集中失败或\n// 集中 errored 在同一生命周期阶段时,追加一条运行级 diagnostic 点名实例、序号区间与阶段。\n// 只指路,不改判定 —— 这里既不读也不写任何 verdict。\nimport type { EvalResult, LifecyclePhase } from \"./types.ts\";\n/** 同一阶段上至少这么多条后承接失败才算「集中」;低于它是零散失败,不发诊断(不误报)。 */\nconst CLUSTER_MIN = 2;\nexport interface ReuseContaminationNotice {\n    /** 承接这些 Attempt 的实验;裸 run(无 experimentId)用 undefined。 */\n    experimentId?: string;\n    /** 本次 Run 内的 Sandbox 编号。 */\n    reuseSandbox: number;\n    /** 集中失败所在的生命周期阶段。 */\n    phase: LifecyclePhase;\n    /** 集中失败的承接序号区间(闭区间)与条数。 */\n    fromOrdinal: number;\n    toOrdinal: number;\n    count: number;\n}\nfunction failing(result: EvalResult): boolean {\n    return result.verdict === \"failed\" || result.verdict === \"errored\";\n}\n/** 失败落在哪个阶段:errored 读 `error.phase`;断言失败没有 error,归 `eval.run`(判定发生地)。 */\nfunction phaseOf(result: EvalResult): LifecyclePhase {\n    return (result.error?.origin.scope === \"attempt\" ? result.error.origin.phase : undefined) ?? \"eval.run\";\n}\n/**\n * 按实例 × 承接序号聚合出复用污染线索。只看声明了复用(`sandbox.reused`)且带完整调度事实的\n * 结果;首承接(序号 1)缺失或本身失败的实例整个跳过——那种失败与「上一条 Attempt 的残留」\n * 无关,报出来就是误报。\n */\nexport function detectReuseContamination(results: readonly EvalResult[]): ReuseContaminationNotice[] {\n    const byInstance = new Map<string, {\n        experimentId?: string;\n        reuseSandbox: number;\n        attempts: EvalResult[];\n    }>();\n    for (const result of results) {\n        const sandbox = result.sandbox;\n        if (!sandbox?.reused || sandbox.reuseSandbox === undefined || sandbox.reuseOrdinal === undefined)\n            continue;\n        const key = `${result.experimentId ?? \"\"}#${sandbox.reuseSandbox}`;\n        let group = byInstance.get(key);\n        if (!group) {\n            group = { ...(result.experimentId !== undefined ? { experimentId: result.experimentId } : {}), reuseSandbox: sandbox.reuseSandbox, attempts: [] };\n            byInstance.set(key, group);\n        }\n        group.attempts.push(result);\n    }\n    const notices: ReuseContaminationNotice[] = [];\n    for (const group of byInstance.values()) {\n        const first = group.attempts.find((r) => r.sandbox!.reuseOrdinal === 1);\n        // 首承接没跑到、或首承接自己就失败:这台实例的失败不能归因到「前一条 Attempt 的残留」。\n        if (!first || failing(first))\n            continue;\n        const laterFailures = group.attempts.filter((r) => r.sandbox!.reuseOrdinal! >= 2 && failing(r));\n        const byPhase = new Map<LifecyclePhase, number[]>();\n        for (const attempt of laterFailures) {\n            const phase = phaseOf(attempt);\n            const ordinals = byPhase.get(phase) ?? [];\n            ordinals.push(attempt.sandbox!.reuseOrdinal!);\n            byPhase.set(phase, ordinals);\n        }\n        for (const [phase, ordinals] of byPhase) {\n            if (ordinals.length < CLUSTER_MIN)\n                continue;\n            const sorted = [...ordinals].sort((a, b) => a - b);\n            notices.push({\n                ...(group.experimentId !== undefined ? { experimentId: group.experimentId } : {}),\n                reuseSandbox: group.reuseSandbox,\n                phase,\n                fromOrdinal: sorted[0]!,\n                toOrdinal: sorted[sorted.length - 1]!,\n                count: sorted.length,\n            });\n        }\n    }\n    return notices;\n}\n/** 诊断正文:点名实例、序号区间与阶段,并说清它只是线索。 */\nexport function reuseContaminationMessage(notice: ReuseContaminationNotice): string {\n    const where = notice.experimentId !== undefined ? ` in experiment \"${notice.experimentId}\"` : \"\";\n    return (`  · [sandbox] reused sandbox #${notice.reuseSandbox}${where} handled its first attempt cleanly, but ` +\n        `${notice.count} later attempts (handoff ${notice.fromOrdinal}-${notice.toOrdinal}) all stopped in ${notice.phase}. ` +\n        \"Files left outside workdir by an earlier attempt are a likely cause: setup must be idempotent and must not \" +\n        \"depend on anything the previous attempt left behind. This only points at a suspect; it changes no verdict.\");\n}\n"]}
+{"version":3,"file":"reuse-diagnostics.js","sourceRoot":"","sources":["reuse-diagnostics.ts"],"names":[],"mappings":";;AAgCA,gEA+CC;AAED,gEAMC;AAhFD,iDAAiD;AACjD,MAAM,WAAW,GAAG,CAAC,CAAC;AAatB,SAAS,OAAO,CAAC,MAAkB;IAC/B,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS,CAAC;AACvE,CAAC;AACD,yEAAyE;AACzE,SAAS,OAAO,CAAC,MAAkB;IAC/B,OAAO,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,UAAU,CAAC;AAC5G,CAAC;AACD;;;GAGG;AACH,SAAgB,0BAA0B,CAAC,OAA8B;IACrE,MAAM,UAAU,GAAG,IAAI,GAAG,EAItB,CAAC;IACL,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC3B,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;QAC/B,IAAI,CAAC,OAAO,EAAE,MAAM,IAAI,OAAO,CAAC,YAAY,KAAK,SAAS,IAAI,OAAO,CAAC,YAAY,KAAK,SAAS;YAC5F,SAAS;QACb,MAAM,GAAG,GAAG,GAAG,MAAM,CAAC,YAAY,IAAI,EAAE,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QACnE,IAAI,KAAK,GAAG,UAAU,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAChC,IAAI,CAAC,KAAK,EAAE,CAAC;YACT,KAAK,GAAG,EAAE,GAAG,CAAC,MAAM,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,MAAM,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,YAAY,EAAE,OAAO,CAAC,YAAY,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;YAClJ,UAAU,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QAC/B,CAAC;QACD,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAChC,CAAC;IACD,MAAM,OAAO,GAAgC,EAAE,CAAC;IAChD,KAAK,MAAM,KAAK,IAAI,UAAU,CAAC,MAAM,EAAE,EAAE,CAAC;QACtC,MAAM,KAAK,GAAG,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAQ,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC;QACxE,wBAAwB;QACxB,IAAI,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC;YACxB,SAAS;QACb,MAAM,aAAa,GAAG,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAQ,CAAC,YAAa,IAAI,CAAC,IAAI,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAChG,MAAM,OAAO,GAAG,IAAI,GAAG,EAA4B,CAAC;QACpD,KAAK,MAAM,OAAO,IAAI,aAAa,EAAE,CAAC;YAClC,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;YAC/B,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;YAC1C,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,OAAQ,CAAC,YAAa,CAAC,CAAC;YAC9C,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;QACjC,CAAC;QACD,KAAK,MAAM,CAAC,KAAK,EAAE,QAAQ,CAAC,IAAI,OAAO,EAAE,CAAC;YACtC,IAAI,QAAQ,CAAC,MAAM,GAAG,WAAW;gBAC7B,SAAS;YACb,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YACnD,OAAO,CAAC,IAAI,CAAC;gBACT,GAAG,CAAC,KAAK,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,KAAK,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACjF,YAAY,EAAE,KAAK,CAAC,YAAY;gBAChC,KAAK;gBACL,WAAW,EAAE,MAAM,CAAC,CAAC,CAAE;gBACvB,SAAS,EAAE,MAAM,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAE;gBACrC,KAAK,EAAE,MAAM,CAAC,MAAM;aACvB,CAAC,CAAC;QACP,CAAC;IACL,CAAC;IACD,OAAO,OAAO,CAAC;AACnB,CAAC;AACD,kCAAkC;AAClC,SAAgB,0BAA0B,CAAC,MAAiC;IACxE,MAAM,KAAK,GAAG,MAAM,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,mBAAmB,MAAM,CAAC,YAAY,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;IACjG,OAAO,CAAC,iCAAiC,MAAM,CAAC,YAAY,GAAG,KAAK,0CAA0C;QAC1G,GAAG,MAAM,CAAC,KAAK,4BAA4B,MAAM,CAAC,WAAW,IAAI,MAAM,CAAC,SAAS,oBAAoB,MAAM,CAAC,KAAK,IAAI;QACrH,gHAAgH;QAChH,0HAA0H,CAAC,CAAC;AACpI,CAAC","sourcesContent":["// 复用失败的收尾诊断(契约见 docs/feature/sandbox/reuse.md「复用失败的可观察性」)。\n//\n// 有意持久状态是合法的实验输入；失败序列本身不能证明污染或根因。\n// Run 收尾按 Sandbox 实例与承接序号聚合:首承接正常、而同一实例序号 ≥ 2 的 Attempt 集中失败或\n// 集中 errored 在同一生命周期阶段时,追加一条运行级 diagnostic 点名实例、序号区间与阶段。\n// 只指路,不改判定 —— 这里既不读也不写任何 verdict。\nimport type { EvalResult, LifecyclePhase } from \"./types.ts\";\n/** 同一阶段上至少这么多条后承接失败才算「集中」;低于它是零散失败,不发诊断(不误报)。 */\nconst CLUSTER_MIN = 2;\nexport interface ReuseFailureClusterNotice {\n    /** 承接这些 Attempt 的实验;裸 run(无 experimentId)用 undefined。 */\n    experimentId?: string;\n    /** 本次 Run 内的 Sandbox 编号。 */\n    reuseSandbox: number;\n    /** 集中失败所在的生命周期阶段。 */\n    phase: LifecyclePhase;\n    /** 集中失败的承接序号区间(闭区间)与条数。 */\n    fromOrdinal: number;\n    toOrdinal: number;\n    count: number;\n}\nfunction failing(result: EvalResult): boolean {\n    return result.verdict === \"failed\" || result.verdict === \"errored\";\n}\n/** 失败落在哪个阶段:errored 读 `error.phase`;断言失败没有 error,归 `eval.run`(判定发生地)。 */\nfunction phaseOf(result: EvalResult): LifecyclePhase {\n    return (result.error?.origin.scope === \"attempt\" ? result.error.origin.phase : undefined) ?? \"eval.run\";\n}\n/**\n * 按实例 × 承接序号聚合复用后的失败。只看声明了复用且带完整调度事实的结果；\n * 首承接缺失或本身失败时没有成功基准，不报告后续失败聚集。\n */\nexport function detectReuseFailureClusters(results: readonly EvalResult[]): ReuseFailureClusterNotice[] {\n    const byInstance = new Map<string, {\n        experimentId?: string;\n        reuseSandbox: number;\n        attempts: EvalResult[];\n    }>();\n    for (const result of results) {\n        const sandbox = result.sandbox;\n        if (!sandbox?.reused || sandbox.reuseSandbox === undefined || sandbox.reuseOrdinal === undefined)\n            continue;\n        const key = `${result.experimentId ?? \"\"}#${sandbox.reuseSandbox}`;\n        let group = byInstance.get(key);\n        if (!group) {\n            group = { ...(result.experimentId !== undefined ? { experimentId: result.experimentId } : {}), reuseSandbox: sandbox.reuseSandbox, attempts: [] };\n            byInstance.set(key, group);\n        }\n        group.attempts.push(result);\n    }\n    const notices: ReuseFailureClusterNotice[] = [];\n    for (const group of byInstance.values()) {\n        const first = group.attempts.find((r) => r.sandbox!.reuseOrdinal === 1);\n        // 首承接缺失或失败时，没有成功基准可供比较。\n        if (!first || failing(first))\n            continue;\n        const laterFailures = group.attempts.filter((r) => r.sandbox!.reuseOrdinal! >= 2 && failing(r));\n        const byPhase = new Map<LifecyclePhase, number[]>();\n        for (const attempt of laterFailures) {\n            const phase = phaseOf(attempt);\n            const ordinals = byPhase.get(phase) ?? [];\n            ordinals.push(attempt.sandbox!.reuseOrdinal!);\n            byPhase.set(phase, ordinals);\n        }\n        for (const [phase, ordinals] of byPhase) {\n            if (ordinals.length < CLUSTER_MIN)\n                continue;\n            const sorted = [...ordinals].sort((a, b) => a - b);\n            notices.push({\n                ...(group.experimentId !== undefined ? { experimentId: group.experimentId } : {}),\n                reuseSandbox: group.reuseSandbox,\n                phase,\n                fromOrdinal: sorted[0]!,\n                toOrdinal: sorted[sorted.length - 1]!,\n                count: sorted.length,\n            });\n        }\n    }\n    return notices;\n}\n/** 诊断正文:点名实例、序号区间与阶段,并说清它只是线索。 */\nexport function reuseFailureClusterMessage(notice: ReuseFailureClusterNotice): string {\n    const where = notice.experimentId !== undefined ? ` in experiment \"${notice.experimentId}\"` : \"\";\n    return (`  · [sandbox] reused sandbox #${notice.reuseSandbox}${where} handled its first attempt cleanly, but ` +\n        `${notice.count} later attempts (handoff ${notice.fromOrdinal}-${notice.toOrdinal}) all stopped in ${notice.phase}. ` +\n        \"This failure pattern does not establish a cause. Intentional persistent memory and checkpoints are supported; \" +\n        \"check expected state, unexpected leftovers, and repeatable setup against the experiment's design. No verdict is changed.\");\n}\n"]}

package/dist/runner/reuse-diagnostics.d.cts

@@ -1,5 +1,5 @@
 import type { EvalResult, LifecyclePhase } from "./types.cjs";
-export interface ReuseContaminationNotice {
+export interface ReuseFailureClusterNotice {
     /** 承接这些 Attempt 的实验;裸 run(无 experimentId)用 undefined。 */
     experimentId?: string;
     /** 本次 Run 内的 Sandbox 编号。 */
@@ -12,10 +12,9 @@ export interface ReuseContaminationNotice {
     count: number;
 }
 /**
- * 按实例 × 承接序号聚合出复用污染线索。只看声明了复用(`sandbox.reused`)且带完整调度事实的
- * 结果;首承接(序号 1)缺失或本身失败的实例整个跳过——那种失败与「上一条 Attempt 的残留」
- * 无关,报出来就是误报。
+ * 按实例 × 承接序号聚合复用后的失败。只看声明了复用且带完整调度事实的结果；
+ * 首承接缺失或本身失败时没有成功基准，不报告后续失败聚集。
  */
-export declare function detectReuseContamination(results: readonly EvalResult[]): ReuseContaminationNotice[];
+export declare function detectReuseFailureClusters(results: readonly EvalResult[]): ReuseFailureClusterNotice[];
 /** 诊断正文:点名实例、序号区间与阶段,并说清它只是线索。 */
-export declare function reuseContaminationMessage(notice: ReuseContaminationNotice): string;
+export declare function reuseFailureClusterMessage(notice: ReuseFailureClusterNotice): string;

package/dist/runner/reuse-diagnostics.d.mts

@@ -1,5 +1,5 @@
 import type { EvalResult, LifecyclePhase } from "./types.mjs";
-export interface ReuseContaminationNotice {
+export interface ReuseFailureClusterNotice {
     /** 承接这些 Attempt 的实验;裸 run(无 experimentId)用 undefined。 */
     experimentId?: string;
     /** 本次 Run 内的 Sandbox 编号。 */
@@ -12,10 +12,9 @@ export interface ReuseContaminationNotice {
     count: number;
 }
 /**
- * 按实例 × 承接序号聚合出复用污染线索。只看声明了复用(`sandbox.reused`)且带完整调度事实的
- * 结果;首承接(序号 1)缺失或本身失败的实例整个跳过——那种失败与「上一条 Attempt 的残留」
- * 无关,报出来就是误报。
+ * 按实例 × 承接序号聚合复用后的失败。只看声明了复用且带完整调度事实的结果；
+ * 首承接缺失或本身失败时没有成功基准，不报告后续失败聚集。
  */
-export declare function detectReuseContamination(results: readonly EvalResult[]): ReuseContaminationNotice[];
+export declare function detectReuseFailureClusters(results: readonly EvalResult[]): ReuseFailureClusterNotice[];
 /** 诊断正文:点名实例、序号区间与阶段,并说清它只是线索。 */
-export declare function reuseContaminationMessage(notice: ReuseContaminationNotice): string;
+export declare function reuseFailureClusterMessage(notice: ReuseFailureClusterNotice): string;

package/dist/runner/reuse-diagnostics.d.ts

@@ -1,5 +1,5 @@
 import type { EvalResult, LifecyclePhase } from "./types.js";
-export interface ReuseContaminationNotice {
+export interface ReuseFailureClusterNotice {
     /** 承接这些 Attempt 的实验;裸 run(无 experimentId)用 undefined。 */
     experimentId?: string;
     /** 本次 Run 内的 Sandbox 编号。 */
@@ -12,10 +12,9 @@ export interface ReuseContaminationNotice {
     count: number;
 }
 /**
- * 按实例 × 承接序号聚合出复用污染线索。只看声明了复用(`sandbox.reused`)且带完整调度事实的
- * 结果;首承接(序号 1)缺失或本身失败的实例整个跳过——那种失败与「上一条 Attempt 的残留」
- * 无关,报出来就是误报。
+ * 按实例 × 承接序号聚合复用后的失败。只看声明了复用且带完整调度事实的结果；
+ * 首承接缺失或本身失败时没有成功基准，不报告后续失败聚集。
  */
-export declare function detectReuseContamination(results: readonly EvalResult[]): ReuseContaminationNotice[];
+export declare function detectReuseFailureClusters(results: readonly EvalResult[]): ReuseFailureClusterNotice[];
 /** 诊断正文:点名实例、序号区间与阶段,并说清它只是线索。 */
-export declare function reuseContaminationMessage(notice: ReuseContaminationNotice): string;
+export declare function reuseFailureClusterMessage(notice: ReuseFailureClusterNotice): string;

package/dist/runner/reuse-diagnostics.js

@@ -1,8 +1,8 @@
 import { createRequire as __niceevalCreateRequire } from "node:module";
 const __niceevalRequire = __niceevalCreateRequire(import.meta.url);
 const __niceevalCanonical = __niceevalRequire("./reuse-diagnostics.cjs");
-const __niceevalExport0 = __niceevalCanonical["detectReuseContamination"];
-export { __niceevalExport0 as detectReuseContamination };
-const __niceevalExport1 = __niceevalCanonical["reuseContaminationMessage"];
-export { __niceevalExport1 as reuseContaminationMessage };
+const __niceevalExport0 = __niceevalCanonical["detectReuseFailureClusters"];
+export { __niceevalExport0 as detectReuseFailureClusters };
+const __niceevalExport1 = __niceevalCanonical["reuseFailureClusterMessage"];
+export { __niceevalExport1 as reuseFailureClusterMessage };
 //# sourceMappingURL=reuse-diagnostics.js.map

package/dist/runner/reuse-plan.cjs

@@ -1,20 +1,29 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.ProjectTargetReusePlanner = exports.EXECUTION_COMPARISON_SOURCE_STATES = exports.EXECUTION_COMPARISON_PROJECTION_STATE = exports.projectTargetPolicyIdentity = exports.PROJECT_TARGET_RECORD_IDENTITY_MAXIMUM_LENGTH = exports.PROJECT_TARGET_INVOCATION_ID_MAXIMUM_LENGTH = exports.PROJECT_TARGET_POLICY_VERSION = exports.PROJECT_TARGET_POLICY_NAME = void 0;
+exports.mergeCurrentAssessmentRuns = mergeCurrentAssessmentRuns;
 exports.recordedAttemptVerdict = recordedAttemptVerdict;
 exports.validateProjectTargetReusePlanInput = validateProjectTargetReusePlanInput;
 exports.planProjectTargetReuse = planProjectTargetReuse;
 exports.planProjectTargetReuseWithoutSources = planProjectTargetReuseWithoutSources;
+exports.assessCurrentTargetSlot = assessCurrentTargetSlot;
+// @concord-file ne-runner-reuse-plan
+// @concord-implements docs/feature/experiments/cache.md
+// @concord-implements docs/feature/experiments/use-case/cache-explicit-rerun.md
+// @concord-implements docs/feature/experiments/use-case/rerun-all.md
 const effect_1 = require("effect");
+const errors_ts_1 = require("../record/reader/errors.cjs");
+const identifiers_ts_1 = require("../record/codec/identifiers.cjs");
+const core_ts_1 = require("../record/codec/core.cjs");
 const score_ts_1 = require("../eval/record/score.cjs");
 const rename_identity_ts_1 = require("./rename-identity.cjs");
 const catalog_ts_1 = require("../record/family/catalog.cjs");
 const eligibility_ts_1 = require("../eval/record/eligibility.cjs");
 const membership_provenance_ts_1 = require("../eval/record/membership-provenance.cjs");
 const verdict_ts_1 = require("../eval/record/verdict.cjs");
-const identifiers_ts_1 = require("../record/model/identifiers.cjs");
+const identifiers_ts_2 = require("../record/model/identifiers.cjs");
 exports.PROJECT_TARGET_POLICY_NAME = "project-target";
-exports.PROJECT_TARGET_POLICY_VERSION = 1;
+exports.PROJECT_TARGET_POLICY_VERSION = 2;
 exports.PROJECT_TARGET_INVOCATION_ID_MAXIMUM_LENGTH = 255;
 exports.PROJECT_TARGET_RECORD_IDENTITY_MAXIMUM_LENGTH = 4096;
 exports.projectTargetPolicyIdentity = Object.freeze({
@@ -33,6 +42,25 @@ exports.EXECUTION_COMPARISON_SOURCE_STATES = [
     ...membership_provenance_ts_1.COMPARISON_SOURCE_STATES,
     exports.EXECUTION_COMPARISON_PROJECTION_STATE.migrationRequired,
 ];
+/** Both consumers retain every expected lifecycle position at one cutoff. */
+function mergeCurrentAssessmentRuns(runs, resources) {
+    return effect_1.Effect.map(effect_1.Effect.forEach(resources, (resource) => {
+        const published = runs.find((run) => run.document.runId === resource.runId);
+        const lifecycle = Object.freeze({ state: resource.state, slots: resource.slots });
+        if (published !== undefined)
+            return effect_1.Effect.succeed(Object.freeze({ ...published, createdRevision: resource.createdRevision, lifecycle }));
+        return effect_1.Schema.decodeUnknownEffect(effect_1.Schema.Struct({
+            runId: identifiers_ts_1.RunIdSchema, experimentId: identifiers_ts_1.ExperimentIdSchema, startedAt: identifiers_ts_1.UtcMillisSchema,
+            expectedSlots: effect_1.Schema.Array(core_ts_1.RecordSlotIdentitySchema),
+        }))({
+            runId: resource.runId, experimentId: resource.experimentId, startedAt: Date.parse(resource.startedAt),
+            expectedSlots: resource.slots.map(({ publication: _publication, ...slot }) => slot),
+        }).pipe(effect_1.Effect.map((document) => Object.freeze({ createdRevision: resource.createdRevision, lifecycle, document, members: Object.freeze([]) })), effect_1.Effect.mapError(() => new errors_ts_1.RecordIntegrityFailure({ code: "record-integrity-failure", runId: resource.runId, reason: "core-invalid" })));
+    }, { concurrency: 1 }), (lifecycle) => Object.freeze([
+        ...lifecycle,
+        ...runs.filter((run) => !resources.some((resource) => resource.runId === run.document.runId)),
+    ]));
+}
 /**
  * Fold the terminal Verdict from the source Attempt's immutable Core outcome
  * and its fixed Assertions facts. This is a projection, not a second durable
@@ -104,6 +132,8 @@ function validateProjectTargetReusePlanInput(input) {
  * Plans only from Record Host selection references. It neither reconstructs
  * attempt handles from strings nor reads private Record paths.
  */
+// @concord-code ne-runner-plan-reuse
+// @concord-implements docs/feature/experiments/cache.md
 function planProjectTargetReuse(input) {
     return effect_1.Effect.suspend(() => {
         const invalid = validateProjectTargetReusePlanInput(input);
@@ -112,15 +142,25 @@ function planProjectTargetReuse(input) {
         return effect_1.Effect.gen(function* () {
             const selection = yield* input.reader.selectRuns({ includePublishedActive: true });
             const readable = yield* effect_1.Effect.forEach(selection.runRefs, (ref) => input.reader.readRun(ref), { concurrency: 1 });
-            const runs = [];
+            const publishedRuns = [];
             const coreIssues = [];
             for (const read of readable) {
                 if (read.state === "available")
-                    runs.push(read.value);
+                    publishedRuns.push(read.value);
                 else if (read.state === "core-invalid")
                     coreIssues.push(...read.issues);
             }
-            const byRunId = new Map(runs.map((run) => [run.document.runId, run]));
+            const resources = [];
+            let afterRunId = "";
+            for (;;) {
+                const page = yield* input.reader.listRunResources({ afterRunId, pageSize: 100 });
+                resources.push(...page.runs);
+                if (page.nextAfterRunId === null)
+                    break;
+                afterRunId = page.nextAfterRunId;
+            }
+            const runs = yield* mergeCurrentAssessmentRuns(publishedRuns, resources);
+            const byRunId = new Map(publishedRuns.map((run) => [run.document.runId, run]));
             const slots = [];
             for (const target of flattenTargetSlots(input.target)) {
                 slots.push(yield* planTargetSlot({
@@ -170,6 +210,22 @@ function planProjectTargetReuseWithoutSources(input) {
 }
 exports.ProjectTargetReusePlanner = Object.freeze({ plan: planProjectTargetReuse });
 function planTargetSlot(input) {
+    return assessCurrentTargetSlot({
+        ...input,
+        readAttempt: (ref) => input.reader.readAttempt(ref).pipe(effect_1.Effect.map((read) => read.state !== "available" ? read : {
+            state: "available",
+            value: { ...read.value, publicationAvailable: read.value.ref.publicationIdentity !== undefined },
+        })),
+        readAssertions: (attempt) => input.reader.read(attempt.owner, catalog_ts_1.NiceEvalRecordAttachments.assertions),
+        readActivities: (attempt) => input.reader.read(attempt.owner, catalog_ts_1.NiceEvalRecordAttachments.runnerActivities.attempt),
+        sourceFor: (attempt, sourceBarrier) => {
+            if (!hasPublicationIdentity(attempt))
+                throw new Error("Expected a published Attempt");
+            return candidateFor({ target: input.target, sourceBarrier, attempt, byRunId: input.byRunId }).source;
+        },
+    });
+}
+function assessCurrentTargetSlot(input) {
     return effect_1.Effect.gen(function* () {
         if (input.target.reuseEligibility !== undefined) {
             return gapSlot(input.target, {
@@ -213,15 +269,18 @@ function planTargetSlot(input) {
         }
         const member = sourceRun.members.find((candidate) => candidate.document.slotId === expected.slotId);
         if (member === undefined || member.attempt === null) {
+            const publication = sourceRun.lifecycle?.slots.find((slot) => slot.slotId === expected.slotId)?.publication;
             return gapSlot(input.target, {
                 reason: "source-member-missing",
                 scope: "slot",
                 issues: [],
                 sourceBarrier,
-                comparisons: [],
+                comparisons: publication?.state === "pending" || publication?.state === "absent"
+                    ? [comparison("core", "attempt-outcome", "unavailable", "not-comparable", publication.state === "pending" ? "pending" : publication.reason)]
+                    : [],
             });
         }
-        const readAttempt = yield* input.reader.readAttempt(member.attempt);
+        const readAttempt = yield* input.readAttempt(member.attempt);
         if (readAttempt.state !== "available") {
             return gapSlot(input.target, {
                 reason: "source-core-invalid",
@@ -231,7 +290,7 @@ function planTargetSlot(input) {
                 comparisons: [],
             });
         }
-        if (!hasPublicationIdentity(readAttempt.value)) {
+        if (!readAttempt.value.publicationAvailable) {
             return gapSlot(input.target, {
                 reason: "source-publication-unavailable",
                 scope: "slot",
@@ -240,13 +299,10 @@ function planTargetSlot(input) {
                 comparisons: [],
             });
         }
-        const candidateResolution = candidateFor({
-            target: input.target,
-            sourceBarrier,
-            attempt: readAttempt.value,
-            byRunId: input.byRunId,
-        });
-        if (candidateResolution === undefined) {
+        const document = readAttempt.value.document;
+        const originRun = input.byRunId.get(document.originRunId);
+        const originExpected = originRun?.document.expectedSlots.find((slot) => slot.slotId === document.slotId);
+        if (originExpected === undefined || document.evalId !== originExpected.evalId || document.executionIdentityDigest !== originExpected.executionIdentityDigest) {
             return gapSlot(input.target, {
                 reason: "source-core-invalid",
                 scope: "slot",
@@ -255,20 +311,21 @@ function planTargetSlot(input) {
                 comparisons: [],
             });
         }
-        const candidate = candidateResolution.source;
-        const ordinaryIdentity = reusableSlotIdentityMatches({ target: input.target, sourceExpected: expected, originExpected: candidateResolution.originExpected });
+        const candidate = input.sourceFor(readAttempt.value, sourceBarrier);
+        const ordinaryIdentity = reusableSlotIdentityMatches({ target: input.target, sourceExpected: expected, originExpected });
         if (!ordinaryIdentity) {
             const originRun = input.byRunId.get(candidate.origin.runId);
             const pureRename = targetMatchesLogicalIdentity(input.target, expected) &&
-                candidateResolution.originExpected.evalId === input.target.evalId &&
-                candidateResolution.originExpected.attemptOrdinal === input.target.attempt &&
-                (0, rename_identity_ts_1.executionDigestForExperiment)(input.target, originRun.document.experimentId) === candidateResolution.originExpected.executionIdentityDigest;
+                originExpected.evalId === input.target.evalId &&
+                originExpected.attemptOrdinal === input.target.attempt &&
+                (0, rename_identity_ts_1.executionDigestForExperiment)(input.target, originRun.document.experimentId) === originExpected.executionIdentityDigest;
             if (!pureRename)
                 return gapSlot(input.target, {
                     reason: "identity-mismatch", scope: "slot", issues: [], sourceBarrier, candidate,
                     comparisons: [identityMismatchComparison()],
                 });
-            if (!(0, rename_identity_ts_1.hasProvenNoExperimentHooks)(sourceRun.document.context) ||
+            if (sourceRun.document.context === undefined || originRun.document.context === undefined ||
+                !(0, rename_identity_ts_1.hasProvenNoExperimentHooks)(sourceRun.document.context) ||
                 !(0, rename_identity_ts_1.hasProvenNoExperimentHooks)(originRun.document.context))
                 return gapSlot(input.target, {
                     reason: "adoption-unproven", scope: "slot", issues: [], sourceBarrier, candidate, comparisons: [],
@@ -289,7 +346,7 @@ function planTargetSlot(input) {
                     reason: "adoption-unproven", scope: "slot", issues: [], sourceBarrier, candidate, comparisons: [],
                 });
         }
-        const assertions = yield* input.reader.read(readAttempt.value.owner, catalog_ts_1.NiceEvalRecordAttachments.assertions);
+        const assertions = yield* input.readAssertions(readAttempt.value);
         if (assertions.state !== "available") {
             const problem = attachmentProblem(assertions);
             return gapSlot(input.target, {
@@ -318,6 +375,14 @@ function planTargetSlot(input) {
                 ],
             });
         }
+        if (input.target.evaluationKind === "score") {
+            const score = (0, score_ts_1.foldRecordedAttemptScore)({ outcome: readAttempt.value.document.outcome, assertions: assertions.value });
+            if (effect_1.Result.isFailure(score) || score.success.state !== "complete")
+                return gapSlot(input.target, {
+                    reason: "score-incomplete", scope: "slot", issues: [], sourceBarrier, candidate,
+                    comparisons: [comparison("niceeval.assertions", "assertion-verdict", "available", "ineligible", "score-incomplete")],
+                });
+        }
         if (verdict !== "passed" && verdict !== "failed") {
             return gapSlot(input.target, {
                 reason: "verdict-ineligible",
@@ -328,15 +393,7 @@ function planTargetSlot(input) {
                 comparisons: [verdictComparison(verdict, "ineligible")],
             });
         }
-        if (!ordinaryIdentity && input.target.evaluationKind === "score") {
-            const score = (0, score_ts_1.foldRecordedAttemptScore)({ outcome: readAttempt.value.document.outcome, assertions: assertions.value });
-            if (effect_1.Result.isFailure(score) || score.success.state !== "complete")
-                return gapSlot(input.target, {
-                    reason: "verdict-ineligible", scope: "slot", issues: [], sourceBarrier, candidate,
-                    comparisons: [verdictComparison(verdict, "ineligible")],
-                });
-        }
-        const activities = yield* input.reader.read(readAttempt.value.owner, catalog_ts_1.NiceEvalRecordAttachments.runnerActivities.attempt);
+        const activities = yield* input.readActivities(readAttempt.value);
         if (activities.state !== "available") {
             const problem = attachmentProblem(activities);
             return gapSlot(input.target, {
@@ -410,20 +467,15 @@ function planTargetSlot(input) {
 }
 function latestSourceRun(runs, target) {
     const candidates = runs.filter((run) => run.document.experimentId === target.experimentId
-        && run.document.expectedSlots.some((slot) => slot.evalId === target.evalId));
+        // Execution refreshes may see their own already-created pending Run.
+        // Logical current targets have no Run identity and keep every barrier.
+        && (!("runId" in target) || run.document.runId !== target.runId)
+        && run.document.expectedSlots.some((slot) => slot.evalId === target.evalId && slot.attemptOrdinal === target.attempt));
     return candidates.sort((left, right) => {
-        const a = left.createdRevision, b = right.createdRevision;
-        if (a !== undefined && b !== undefined)
-            return a - b || (0, identifiers_ts_1.compareCanonicalIdentity)(left.document.runId, right.document.runId);
-        return compareReadableRuns(left, right);
+        const a = left.createdRevision ?? 0, b = right.createdRevision ?? 0;
+        return a - b || (0, identifiers_ts_2.compareCanonicalIdentity)(left.document.runId, right.document.runId);
     }).at(-1);
 }
-function compareReadableRuns(left, right) {
-    if (left.document.startedAt !== right.document.startedAt) {
-        return left.document.startedAt < right.document.startedAt ? -1 : 1;
-    }
-    return (0, identifiers_ts_1.compareCanonicalIdentity)(left.document.runId, right.document.runId);
-}
 function hasPublicationIdentity(attempt) {
     return attempt.ref.publicationIdentity !== undefined;
 }
@@ -463,7 +515,7 @@ function targetMatchesLogicalIdentity(target, slot) {
         && target.executionIdentityDigest === slot.executionIdentityDigest;
 }
 function targetMatchesSlotIdentity(target, slot) {
-    return target.slotId === slot.slotId
+    return (!("slotId" in target) || target.slotId === slot.slotId)
         && target.evalId === slot.evalId
         && target.attempt === slot.attemptOrdinal
         && target.executionIdentityDigest === slot.executionIdentityDigest;
@@ -573,11 +625,11 @@ function isTargetRun(value) {
         return false;
     const run = value;
     return (typeof run.runId === "string"
-        && (0, identifiers_ts_1.isPortableSegment)(run.runId)
+        && (0, identifiers_ts_2.isPortableSegment)(run.runId)
         && typeof run.experimentId === "string"
         && isBoundedRecordIdentity(run.experimentId)
         && typeof run.startedAt === "number"
-        && (0, identifiers_ts_1.isUtcMillis)(run.startedAt)
+        && (0, identifiers_ts_2.isUtcMillis)(run.startedAt)
         && Array.isArray(run.slots));
 }
 function isTargetSlot(value) {
@@ -585,9 +637,9 @@ function isTargetSlot(value) {
         return false;
     const slot = value;
     return (typeof slot.runId === "string"
-        && (0, identifiers_ts_1.isPortableSegment)(slot.runId)
+        && (0, identifiers_ts_2.isPortableSegment)(slot.runId)
         && typeof slot.slotId === "string"
-        && (0, identifiers_ts_1.isPortableSegment)(slot.slotId)
+        && (0, identifiers_ts_2.isPortableSegment)(slot.slotId)
         && typeof slot.experimentId === "string"
         && isBoundedRecordIdentity(slot.experimentId)
         && typeof slot.evalId === "string"
@@ -599,7 +651,7 @@ function isTargetSlot(value) {
         && (slot.reuseEligibility === undefined
             || slot.reuseEligibility === "adapter-behavior-revision-required")
         && typeof slot.executionIdentityDigest === "string"
-        && (0, identifiers_ts_1.isSha256Digest)(slot.executionIdentityDigest)
+        && (0, identifiers_ts_2.isSha256Digest)(slot.executionIdentityDigest)
         && isExecutionIdentity(slot.inputIdentity)
         && isExecutionIdentity(slot.configIdentity)
         && (slot.timeout === undefined || isExecutionDurationLimit(slot.timeout)));