[wip ai slop] Verify immutable receipts before ingest / 导入前验证不可变回执

[functionstackx]

Current status: Production was rolled back at the user’s request to the pre-#1179 deployment at 9bb7b13eb4985217a6282f340459fd5948613276 (Vercel). The exact code revert is #1180, merged with explicit user approval at 92fef485edd5ae61fe49d01f0e41b67492263bee. Current master exactly matches the pre-#1179 tree. All other PRs remain unmerged. INFX_PHASE1_READER_REVISION has been removed; Phase 1 publication remains gated. The additive table and migration ledger are preserved. Deployment and migration results below are historical evidence, not current reader readiness.

Native Phase 1 measurement artifacts now require an independently issued receipt before staging reset or database writes. Artifact downloads use exact API IDs, stream ZIPs to disk, and validate bounded member streams before safe extraction; imports bind an immutable source run/attempt and support resuming only the same accepted snapshot. Production additionally requires a publication record tied to the executing ingest SHA, including when source and merge run IDs are equal.

Receipt ingestion selects only bound normalized benchmark members, imports bound eval samples from both aggregate and raw layouts, and rejects run-mode or purge skips before completing a snapshot. Same-receipt replay verifies successful per-file persistence. Aggregate TP8/TP4 evals retain zero prefill/decode workers and strict-filter sample counts. A post-refresh verifier checks source-specific and latest curves, trace detail and eval visibility. Migration 016_measurement_snapshots.sql stores accepted source receipt identity; the later publication record remains a separately retained artifact. Legacy input behavior remains covered.

A dedicated migrate-database.yml workflow now provides an explicit staging/production migration-only route after landing on master. It binds dispatch to the exact reviewed app SHA, selects only the target writer credential, uses the existing migration command, and verifies migration 016, required column types/nullability and source/run/attempt primary key in a read-only transaction. It retains a credential-free JSON readiness artifact and performs no unrelated ingest, reset or cache refresh. Keep native receipts gated until production deployment and schema verification both succeed.

Local validation for 9003c0bcc2a46d52167ca007d50dd26cc0c911ac: all 6,892 unit tests passed (four skipped), including six new real PGlite preservation/replay/schema-drift cases; smoke tests passed with 56 component and 128 integration tests. Typecheck, targeted lint/format, actionlint, the full repository-configured Zizmor audit and the authenticated new-workflow audit passed. Fresh GitHub CI passed at this head: 6,892 unit tests and typecheck (four skipped), 486 component tests and 1,033 integration tests in each of Chrome and Firefox, lint/format, repository-configured Zizmor and CodeQL. Bugbot completed successfully with zero unresolved review threads.

This reader prerequisite was normally squash-merged to master as 481a8622cc9bc27feae775850e241ec967bac1e3; verify its production deployment and both target migration reports before enabling the H100 producer. Source collector and producer changes are SemiAnalysisAI/InferenceX#3298 and SemiAnalysisAI/InferenceX#3299; the native prerequisite is SemiAnalysisAI/srt-slurm#21. Set the issuer SHA/workflow allowlist and use the final deployed app SHA in reader/ingest/publication pins. The exact merged app SHA passed both real migration-only runs: staging 35478033492 and production 35478066182. Retained JSON reports verify migration 016, required column types/nullability and the source/run/attempt primary key. The exact merged SHA also has a successful Vercel Production deployment, GitHub deployment 6547210249, completed at 2026-09-20T00:11:38Z. No native receipt ingestion or real pilot measurement import has been performed.

Additional local real database replay/failure tests and a 256 MiB archive extraction below 192 MiB peak RSS passed. The prior three Bugbot findings and the additional receipt-completeness gap are fixed. A separate stronger auditor persona on the changed workflows retains 19 repository-secret advisories (15 baseline, two receipt token uses and two migration writer uses); no suppressions or credential-policy changes were added. Hardware and actual publication evidence remain pending. This changes official ingestion contracts, not unofficial overlay presentation.

AI model disclosure

Codex is described by this runtime as based on GPT-6; the exact model/version identifier was not exposed and could not be verified. The primary agent integrated the change; delegated agents using the inherited runtime implemented and reviewed the publication readers, Python producer contracts, migration readiness workflow/schema verification and workflow audits.

中文说明

当前状态： 按用户要求，生产环境已回滚到 #1179 合入前的 9bb7b13eb4985217a6282f340459fd5948613276 部署。完整代码撤销见 #1180；经用户明确批准，该撤销 PR 已合入为 92fef485edd5ae61fe49d01f0e41b67492263bee，当前 master 的代码树与 #1179 合入前完全一致。其他 PR 均未合并。INFX_PHASE1_READER_REVISION 已删除，Phase 1 发布继续受限。保留新增数据库表及迁移记录。下方部署和迁移结果仅作为历史证据，不表示当前 reader 已就绪。

原生阶段 1 的测量产物在重置 staging 或写数据库前，必须通过独立签发的回执。下载使用准确的 GitHub API artifact ID，将 ZIP 流式写入磁盘，并在安全解包前通过有界成员流完成校验；导入绑定不可变的 source run/attempt，部分失败只允许续传同一已接受快照。生产导入还必须提供与实际 ingest SHA 一致的发布记录，即使 source 与 merge run ID 相同也不能省略。

回执导入仅选择绑定的标准化 benchmark 成员，支持 aggregate 与 raw 两种布局中的绑定 eval 样本，并在快照完成前拒绝运行模式或 purge 导致的跳点。同一回执重放验证每个文件的实际成功写入。聚合 TP8/TP4 eval 保留零 prefill/decode worker，并验证 strict filter 样本数。刷新后的验证器检查源 run 与最新曲线、trace 详情及 eval 可见性。Migration 016_measurement_snapshots.sql 保存已接受源回执身份；后续发布记录作为独立 artifact 保留。旧输入路径仍有回归覆盖。

新增 migrate-database.yml，合入 master 后可显式选择 staging 或 production 独立执行迁移。调度绑定准确的已审查 app SHA，仅提供目标数据库的 writer 凭证，沿用现有迁移命令，并在只读事务中确认 migration 016、必需列类型/非空约束及 source/run/attempt 主键。流程保留不含凭证的 JSON 就绪产物，不额外导入结果、不重置数据库，也不刷新缓存。production 部署与 schema 验证都成功之前，继续限制原生回执导入。

9003c0bcc2a46d52167ca007d50dd26cc0c911ac 的本地验证：6,892 项单元测试通过（4 项跳过），包括新增的 6 项真实 PGlite 数据保留/重复执行/schema 不一致测试；smoke 测试的 56 项 component 与 128 项 integration 全部通过。Typecheck、针对性 lint/格式、actionlint、全仓库配置的 Zizmor 及新 workflow 的认证在线审计均通过。该提交的新一轮 GitHub CI 已通过：6,892 项单元测试与 typecheck（4 项跳过）、486 项 component 测试与 Chrome/Firefox 各 1,033 项 integration 测试、lint/格式、仓库配置的 Zizmor 与 CodeQL。Bugbot 审阅成功完成，未解决审阅线程为零。

此 reader 前置能力已正常 squash 合入 master，提交为 481a8622cc9bc27feae775850e241ec967bac1e3；启用 H100 producer 前，应验证该准确提交的 production 部署和两个目标数据库的迁移报告。InferenceX 的 collector 与 producer 分别见 SemiAnalysisAI/InferenceX#3298 和 https://github.com/SemiAnalysisAI/InferenceX/pull/3299；原生前置能力见 https://github.com/SemiAnalysisAI/srt-slurm/pull/21。配置 issuer SHA/workflow allowlist，并使用最终部署 app SHA 作为 reader/ingest/publication pin。准确的合入 app SHA 已通过两次真实独立迁移：staging 35478033492 与 production 35478066182。保留的 JSON 报告确认 migration 016、必需列类型/非空约束及 source/run/attempt 主键。准确的合入 SHA 也已成功完成 Vercel Production 部署，GitHub deployment 为 6547210249，完成时间为 2026-09-20T00:11:38Z。尚未导入原生回执或真实试点测量。

额外的本地真实数据库重放/失败测试，以及峰值 RSS 低于 192 MiB 的 256 MiB 归档解包检查通过。此前三项 Bugbot 发现与额外的回执完整性问题均已修复。对本次改动的 workflow 使用更严格 auditor persona 时保留 19 项仓库 secret 提示（15 项基线、2 项回执 token 引用及 2 项迁移 writer 引用），未添加抑制或改变凭证策略。硬件与实际发布证据仍待补；本变更涉及正式导入契约，不改变 unofficial overlay 展示。

AI 模型披露：运行时将 Codex 描述为基于 GPT-6，但未暴露可核实的精确模型/版本。主代理负责集成；继承相同运行时的子代理负责发布 reader、Python producer 契约、迁移就绪 workflow/schema 验证及 workflow 的实现与审阅。

---

Note

High Risk
Changes official ingest and publication contracts, adds production DB migration and credential-backed cross-repo artifact verification on the critical data path.

Overview
Introduces immutable measurement publication for native Phase 1 ingest: staging and production workflows now verify receipt (and, for production, publication) transport via pinned issuer SHAs before artifact download or DB writes, with new dispatch inputs and post-ingest verify-measurement-publication artifacts.

The db package gains receipt parsing/validation, streaming ZIP extraction with digest checks, measurement_snapshots (migration 016), receipt-bound ingest that resumes partial imports on the same snapshot, and blocks legacy bypass once a run is accepted. Production ingest requires a publication record whose ingest_sha matches the workflow checkout. A new migrate-database workflow applies migrations and emits schema verification reports. Eval mapping treats explicit single-node aggregate deployment as one serving role; eval sample dedup is centralized via projectEvalSamples.

^{Reviewed by Cursor Bugbot for commit 9003c0b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Sep 20, 2026 12:06am UTC

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 5e2b92b. Configure here.}