Simplify reusable sweep ingestion with direct source dispatch / 通过直接分发源运行简化复用扫描摄取

[cquil11]

Summary

• remove the reuse-ingest-artifacts relay job
• send only source-run-id and merge-run-id to InferenceX-app
• reduce manual recovery to the same two-ID dispatch
• keep artifact selection and ingestion entirely in InferenceX-app
• update recovery documentation to match the simple normal-ingestion behavior

End-to-end flow

1. The existing reuse gate authorizes reuse and selects the source PR sweep.
2. InferenceX sends the source run ID and current main publication run ID.
3. InferenceX-app downloads the source artifacts, keeps the newest upload for each exact name, and substitutes the merge-run changelog.
4. The existing database ingestion code handles all benchmark/eval artifact semantics and skips failed benchmark rows.

Run attempts are not sent across repositories because GitHub can derive them from the run IDs.

Validation

• all 29 workflow-gating tests passed
• paired app implementation passed 362 database unit tests, typecheck, lint, and formatting
• live dry runs passed for PR [Agentx][sglang] B200+B300 DeepSeek V4 AgentX config update #2145 source run 29367944489 and PR [AgentX] vLLM DeepSeek-V4 B200 aggregate / vLLM DeepSeek-V4 B200 聚合配置 #2224 source run 29445892486

Related consumer PR: SemiAnalysisAI/InferenceX-app#582

中文说明

• 删除 reuse-ingest-artifacts 中转任务
• 仅向 InferenceX-app 发送 source-run-id 和 merge-run-id
• 手动恢复也简化为相同的双运行 ID 分发
• 产物选择与数据库摄取全部由 InferenceX-app 负责
• 更新恢复文档，使其与简化后的常规摄取行为一致

端到端流程现在是：现有复用门禁选择源 PR 扫描，InferenceX 发送源运行和主分支发布运行 ID，InferenceX-app 下载源产物、按精确名称保留最新上传并替换合并运行变更日志，最后由现有数据库摄取逻辑处理所有产物语义并跳过失败结果。

验证结果：29 个工作流门禁测试全部通过；配套 app 实现通过 362 个数据库单元测试、类型检查、lint 和格式检查；PR #2145 与 PR #2224 的实时 dry run 均成功。

[claude]

Claude finished @cquil11's task in 4m 50s —— View job

---

Review

• Gather context: diff against origin/main, read changed workflow files and tests
• Review for bugs, breaking changes, and contract issues with InferenceX-app
• Run cited tests (blocked by sandbox permissions — verified statically instead, see note)
• Post review feedback

LGTM - no blocking issues found.

I verified the changes hold together:

• The removed relay job's fail-closed role (producer-side validation gating the dispatch) is intentionally moved to the app, and the new trigger-ingest/trigger-agentic-ingest conditions (needs.setup.outputs.reuse-enabled == 'true' replacing needs.reuse-ingest-artifacts.result == 'success') are otherwise equivalent to the old gating.
• The ternary-style payload expression (reuse-enabled == 'true' && reuse-source-run-id || github.run_id) is safe: find_reusable_sweep_run.py always emits a non-empty reuse-source-run-id when reuse-enabled is true (utils/find_reusable_sweep_run.py:488-496), so the || fallback can't fire spuriously on the reuse path.
• upload-changelog-metadata still runs on push-to-main reuse runs (always() && setup.result == 'success'), so the merge run always owns the changelog artifact the app now fetches via merge-run-id.
• The new structural tests' assumptions (single dispatch step at steps[0], exactly two recovery inputs, one recovery job, no attempt fields) all match the actual workflow files.
• Cross-repo payload rename (run-id/run-attempt → source-run-id/merge-run-id) is a breaking contract change, but the rollout ordering in the PR description (app PR [nv] add H200 SGLang disagg configs from srtslurm #582 first, app stays compatible with the old payload until then) handles it.

One non-blocking observation: .claude/commands/recover-failed-ingest.md (lines 114, 323, 452) and KLAUD_DEBUG.md:209 still reference the removed reuse-ingest-artifacts job and producer-side validate_reusable_sweep_artifacts.py step — the skill's acceptance checklist requires "pass reuse-ingest-artifacts consistency validation", which no longer exists in the pipeline. Worth a follow-up so the recovery runbook doesn't point agents at a nonexistent job. Fix this →

Note on validation: the CI sandbox denied running pytest and actionlint directly, so I verified every assertion in the new tests by inspecting the workflow files (job removal, needs lists, dispatch payload keys, recovery inputs). The test-changelog-gate.yml workflow will run the gating suite in CI regardless.