fix(bench): accept expanded ingest run names in history collection - #2322

Summary

Publisher run 42168067244 rejected a successful benchmark ingest because GitHub returned name = "e2e-bench ingest 42158962746-1", not the workflow's static name, "End-to-end benchmark ingest".

The rejected ingest run was dispatched by the Pika GitHub App, the automation identity that submits benchmark results. It was a calibration run, which establishes reference measurements for later benchmark comparisons.

PR, main, nightly and calibration all use this ingest workflow. Collection checks the workflow's identity before reading the tier, so every tier would fail. Calibration verification repeated the same comparison and would fail later in publication.

Changes

Both validators now use the exact repository and workflow path for identity, without comparing display names. The fixtures now match GitHub's expanded run-name, and the documentation no longer claims that the collector checks the run name.

Testing

The fixtures reproduced both failures before the fixes. After the fixes:

• uv run --no-project python -m unittest scripts.tests.test_benchmark_history scripts.tests.test_e2e_calibration scripts.tests.test_e2e_publication ran 32 tests: 31 passed and one existing ELF-tool-dependent test was skipped.
• The collector accepted the real ingest attempt. Calibration verification used real GitHub metadata and artifact bytes, but a local list of calibration references instead of the published list. All 17 benchmark cases found a matching reference. Live publication has not been replayed.

The existing benchmark-contract CI workflow already runs the publication and calibration suites. No workflow change was needed.

Security Considerations

The ingest must still be App-authored, run successfully on main through workflow_dispatch, and match the exact repository, head repository, workflow path and attempt. Its result artifact must match the ingest run and attempt.

The ingest.json receipt still binds the ingest to the workflow run that prepared its benchmark inputs. Both validators retain their artifact-ID, SHA-256 and receipt checks.

Breaking Changes

There are no public API, proof-format or serialization changes.

Replay after merge

Once this reaches main, replay publication without rerunning the benchmarks:

gh workflow run prover-history.yml --repo a16z/jolt --ref main -f run_id=42168032225 -f run_attempt=1

The referenced artifacts must still be retained.