[codex] Recover PR 1798 ingest without rerunning sweep

[Oseltamivir]

What changed

• adds a manually dispatched, CPU-only recovery workflow for PR [NV]Add GLM-5 NVFP4 GB300 disagg-non-mtp TRT-LLM benchmarks via Dynamo  #1798
• validates failed push run 27712344914 and setup job 81976315082
• validates and reuses artifacts from successful source run 27622347964, attempt 2
• reconstructs PR [NV]Add GLM-5 NVFP4 GB300 disagg-non-mtp TRT-LLM benchmarks via Dynamo  #1798 merge metadata without treating its unrelated PR [NVIDIA][GB300] update DSR1 FP8 GB300 TRTLLM image to latest #1767 indentation repair as a changelog deletion
• dispatches the resulting carrier run to InferenceX-app

Root cause

The push-to-main run failed in setup because PR #1798 corrected the malformed PR #1767 changelog indentation. The changelog processor rejects non-whitespace deletions, so artifact reuse and ingest never ran.

Safety

• commit and merge message include [skip-sweep]
• workflow requires an explicit recover-pr-1798 confirmation input
• failed target run/job, source run attempt, PR commit, workflow path, conclusion, and required artifacts are validated
• the reconstructed matrix is validated against the downloaded artifacts before dispatch
• no benchmark matrix or GPU runner is referenced

Validation

• actionlint .github/workflows/recover-pr-1798-ingest.yml
• target/source guard script passed locally
• reusable artifact validator passed: 20 benchmark rows and 2 multinode eval jobs

---

Note

Medium Risk
Touches production ingest via repository dispatch and uses a one-off changelog/git reconstruction; mitigated by hard-coded run/SHA checks, artifact validation, and no GPU matrix.

Overview
Adds a manually dispatched GitHub Actions workflow (recover-pr-1798-ingest.yml) to finish database ingest for PR #1798 after the post-merge run-sweep push failed in setup, without rerunning benchmarks on GPU.

The workflow mirrors the existing PR #1767 recovery pattern but pins specific failed push run/job IDs and a successful PR sweep run (attempt 2), and requires typing recover-pr-1798 to run. It validates those runs via the GitHub API (event, workflow path, SHAs, conclusions, non-expired results_bmk / eval_results_all / run-stats artifacts).

It reconstructs the sweep config by building a synthetic tree from the merge commit: a perl edit on perf-changelog.yaml undoes the unrelated PR #1767 indentation fix so process_changelog.py does not treat it as a forbidden deletion, then runs the validator and uploads artifacts plus changelog metadata. Finally it dispatches ingest-results to InferenceX-app with the recovery run id.

^{Reviewed by Cursor Bugbot for commit 9f0a805. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Oseltamivir]

Recovery validation is complete:

• failed target: run 27712344914, setup job 81976315082
• reusable source: run 27622347964, attempt 2, commit ffe21af3
• reconstructed PR [NV]Add GLM-5 NVFP4 GB300 disagg-non-mtp TRT-LLM benchmarks via Dynamo  #1798 matrix: 20 benchmark rows and 2 multinode eval jobs
• actionlint and the repository artifact validator both pass
• recovery workflow is CPU-only and does not reference benchmark/GPU runners

After this lands on main, I will dispatch it with recover-pr-1798 and verify the downstream InferenceX-app ingest.