Repository navigation
[PR] test: align host test pins and kernel.json snapshots after #5860 phantom pressure guard (fix deterministic red in master | test: 对齐 #5860 幻影压力守卫后的宿主测试钉与 kernel.json 快照(修 master 定时矩阵确定性红) - #5878
Conversation
…ted-state shape #5860 added compress.requireUsageSampleForPressure; the resolved Config snapshot persisted in session.metadata["effectiveConfig"] (src/session.ts cloneSafe) grew by exactly that field, so the metadata section digest drifts deterministically on every lane (F1 of the #5860 born-red wave, signatures pinned in #4599). Regenerated all four snapshots at 5af5f9f with scripts/update-storage-render-goldens.ts; state/blockContents/stats sections unchanged, no behavior change - snapshot realignment only. Verified green under the CI preload env (BILI_MESSAGE_TIMESTAMMS=0 via tests/preload-test-env.mjs).
…ract kernel.json buildKernelGolden() serializes behavior.defaultConfig100k = defaultConfig(100_000); #5860's additive NudgeConfig field broke the canonicalized snapshot (F2 of the #5860 born-red wave). Regenerated with scripts/update-wire-contract-goldens.ts: diff is the single additive field requireUsageSampleForPressure: true in the nudge section - the ledger grows, nothing shrinks (Kernel Boundary rule 4).
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.8514-pr.5878.8a34be4This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr5878.tgz
npm install -g packageInstall instructions are refreshed on each push. |
…orphaned #5860 defers pressure-band injection whenever the token reading is estimate-grade and no billing sample landed this round (documented design, default on; escape hatch compress.requireUsageSampleForPressure = false). Eight host-side test pins assert ESTIMATE-driven injection and now fail deterministically — their upstream mocks never send usage, so every reading in them is estimate-grade: - silent-backend-nudge.test.ts #728A, #728C (precondition), #1119 - fork-nudge-trigger.test.ts anon zero-baseline fork, #1137 image-term fork - issue4241-judge-render-baseline.test.ts teeth assume the engine-appended trailing nudge occupies exitCores.at(-1); the defer shifts the index so the faulted drop no longer hits the owned tail row Each suite predates the guard and pins PRE-GUARD mechanisms (fork baseline regime, image-term arithmetic, judge reconciliation against the rendered baseline) — not sample gating. Rewriting their assertions to post-guard semantics would make them vacuous on a permanently-silent backend (the defer repeats every estimate round; the next usage-grade sample the design waits for never arrives). They therefore opt out through the documented escape hatch, keeping their original coverage byte-for-byte. The guard itself keeps dedicated coverage in tests/nudge-pressure-no-sample-armed.test.ts (run unmodified: green). Local proof at 5af5f9f (fresh npm ci + build, CI-parity env pins): before: 3+3+1 deterministic failures across the three files; after: 11/11 pass incl. the guard suite.
808efb1 to
71a5338
Compare
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。 本 PR 是关闭 #5876 的唯一 PR。与哨兵收敛 PR #5879(ci-fixes→master)在 kernel.json 上有一处字节级相同的单行 hunk——已按 #4599 收敛政策处理(见下)。与 #5868(storage-render 重对齐,不同根因 #5772)、GitHub PR #5859/镜像 #5861(F3 决策线程)相关但机制不同,不并入。 标签: bug + scope/ci + p2(测试对齐修定时矩阵确定性红,无用户面行为变更; 八问
本地证据(双向验证,均为定向套件非全量)
我做的协调动作(均已实际执行并验证)
判定内容已验证可合;合入等 owner 对 F3 语义裁决(镜像 #5861 floor 4 挂着 A/B/C + 默认行为,尚无拍板)。推荐把「接受 D(本 PR 测试侧逃生阀)+ 饥饿问题留在该决策线程继续评估 B」作为一个完整选项:这样 born-red 波(F1/F2/F3)一次清完、master 矩阵全绿,而唯一能真正修复功能性回归的 B(kernel 有界推迟,human-gated)不被抢跑也不被搁置。若拍 B,本 PR 的 6 行逃生阀按有界契约小改重对齐即可(follow-up)。spawn 型集成测试 ~15s 超时偶发为已知 flake,另行跟进,不阻塞。我不执行合并(human-only):PR #5878。 |
(ai: vllm-qwen/qwen3.8-27b) 补一条状态确认(完整审核八问已在上方 13:03Z 回复中发出):本轮审核与协调工作已全部完成——PR 分支已按 #4599 收敛政策重建并推送(移除与 ci-fixes 字节重复的自研 golden 提交,改 merge ci-fixes 带入两族 golden 重对齐),新 head 71a5338,构建管线已在该 head 上触发;本地双向验证通过(master@5af5f9fe 逐名复现 7 fail / 最终树六套件 21/21 pass),标签 bug+scope/ci+p2(+scheduled-red)已就位。当前状态:等 CI 在新 head 上复跑确认 + 等 owner 对 F3 语义的裁决(镜像 #5861 floor 4 🔶 需要决策;推荐选项见上条回复)。无其他待办。 |
|
(ai: sglang-gpu0/qwen3.8-27b) 本 PR(#5860 守卫后的宿主钉/kernel.json 对齐)被 PR #5882 已取代——#5882 已合入 master,且处理得更好:
|
一句话
#5860(幻影压力守卫,kernel 0.0.184)合入时漏了两样东西:host 侧 6 个断言旧语义的测试钉、wire-contract/storage-render 的 golden 快照。本 PR 补这两样——不改任何产品行为。Closes #5876。
分支构成(2026-11-07 review 后收敛)
5bcd6c35b771a5338f92(merge ci-fixes)e7b0cee859storage-render 四线快照重对齐(+#5860 持久化形状)、6436b23b3ckernel.json defaultConfig 块纯增量一行"requireUsageSampleForPressure": true——无行为常量变化、无账本缩减review 说明:原自研 golden 提交与 ci-fixes 上的
6436b23b3c逐字节重复(master 级修复必须收敛到 ci-fixes,#4599 政策),已移除;golden 内容经 merge 原样保留。推送后本分支同时覆盖 born-red 波的 F1(storage-render)+F2(kernel.json)+F3(六条测试钉)。为什么是「加逃生阀」而不是「改写断言」
#5860 的设计(默认开):本轮 token 读数是本地估算且处于压力带 → 注压延后一轮等真实 usage。这 6 个钉全部跑在永不回 usage 的 silent mock upstream 上——新语义下它们等待的那份 usage 永远不来,延后每轮重复,把断言改写成后守卫语义等于把它们变成永不动作的空测。这些套件守的是守卫之前的机制(fork 基线制度 / 图片项计量 / judge 对渲染基线的核对);守卫自己有独立专属套件 tests/nudge-pressure-no-sample-armed.test.ts,本 PR 不碰它。用文档化逃生阀恢复各套件的原始覆盖,是既有配置面里最窄的对齐方式(不新增任何配置形状)。
#4241 的 teeth 单独说明:它假设引擎追加的尾部 nudge 占 exitCores.at(-1);延后让 nudge 缺席后索引整体前移,故障注入恰好砍进设计豁免区(prefix),teeth 哑火——不是 judge 坏了。
证据链
与相邻工作的关系