fix(etl): skip rows InferenceX quarantined and refuse unknown result versions

[adibarra]

Companion to SemiAnalysisAI/InferenceX#3806 and #1293.

InferenceX#3806 makes its collectors leave rows that break the result row contract out of results_<prefix> and eval_results_<prefix>, list them in rejected_rows_* artifacts, and stamp published rows with result_schema_version: 1. This ingest also reads the per-config bmk_* and eval_* artifacts, which still contain the quarantined rows, so those rows could reach the database.

Changes

• Quarantine. The CI ingest (ingest-ci-run.ts) and the GCS restore (ingest-gcs-backup.ts) read rejected_rows_<prefix> and rejected_rows_eval_<prefix> before any write. Both download paths already fetch every artifact, so prepare-ci-artifacts.ts is unchanged.
  • A per-config benchmark file is skipped when its <artifact>/<file> path equals an entry's source and its parsed content equals the entry's row. The content check keeps a list left by an earlier attempt from hiding a file that a later attempt re-uploaded under the same name.
  • A per-config eval task is skipped when its results file and task match the entry's row.source and row.task. Eval sources start with the collector's download directory (eval_results/), so only the last two segments, <artifact>/<file>, are compared.
  • A list that is not standard JSON, or whose entries lack these identifiers, fails the CI ingest before any write and skips that run in a GCS restore. rejected_rows_run_stats is not read because run-stats already omits those entries.
  • A quarantined 8K/1K result is recorded as a PowerX publication error, like a failed one.
• Versions. mapBenchmarkRow, mapAggEvalRow, and a new shared mapRunStats refuse any result_schema_version other than 1. Unstamped rows from current InferenceX main, historical artifacts, and GCS backfills ingest as before.
• Reporting. New skip-tracker counters print in the ingest summary as unsupported result_schema_version and quarantined by InferenceX.
• Docs. docs/data-pipeline.md describes the contract handling.

InferenceX change

Python wrote non-finite floats in rejected_rows.json as bare NaN or Infinity tokens, which JSON.parse rejects, and a NaN row is one of the contract's own rejection cases. SemiAnalysisAI/InferenceX@72e9b2334 on #3806 writes them as strings such as "NaN".

#1293 keeps result_schema_version: 1 out of benchmark metrics. It is independent of this PR, and both are needed.

Checks

bun run typecheck, bun run lint, bun run fmt, and bun run test:unit pass locally.

[vercel]

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

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 8, 2026 3:20pm UTC