fix(db): build each chart line from a single run, no cross-run/date stitching

[adibarra]

Note

Medium Risk
Changes core benchmark read paths and chart semantics (lines may lose concurrencies after partial re-sweeps); matview behavior lags until migration 007 is applied in prod.

Overview
Fixes “frankenlines” where one chart line mixed concurrencies from different workflow runs when a newer partial re-sweep skipped some concurrencies.

getLatestBenchmarks now treats a line as (config_id, benchmark_type, isl, osl): pick the newest run (date, then run_started_at, then workflow_run_id) and return only that run’s concurrencies—partial re-sweeps shorten the line instead of backfilling from older runs. The date-filtered SQL path uses a winners CTE + join instead of DISTINCT ON (…, conc, …); json-provider mirrors the same logic for dump/test mode.

Migration 007 rebuilds the latest_benchmarks matview with the same rule and adds benchmark_type to the unique index. Until prod runs this migration, the no-date “latest” path can still serve the old matview while the date path is already fixed.

New unit tests cover truncation, per-sequence independence, same-day tiebreaks, and as-of-run behavior.

^{Reviewed by Cursor Bugbot for commit 391102d. 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 (UTC)
inferencemax-app	Ready	Preview, Comment	Jun 23, 2026 11:23pm