Re-run failed tests from the PR dashboard

[tarekziade]

Re-run failed tests from the PR dashboard: a picker for failed tests, an action API on the trace-exporter, and targeted CPU/GPU re-run workflows.

• Picker UI. The read-only preview, already deployed, appears in the PR and Branch modals, only for open PRs.
• Action API + durable actions (rerun_actions.py). Actions are stored in SQLite on the run-store volume: type, user, subject, idempotency key, state and the workflow runs they started. A page reload finds an action that is still in flight.
• Workflows. rerun-failed-cpu.yml / rerun-failed-gpu.yml re-run only the selected node IDs. Callers in transformers dispatch them, under workflow-templates/rerun-failed/. The correlation ID in the run-name lets the exporter find the run it dispatched.
• GitHub writes go through serge. Reads use the exporter's own token. can_write, cancelling the PR's running CI and dispatching the re-run are sent to serge's /dashboard/* API as the acting user (Add /dashboard API for the dashboard's rerun GitHub writes serge#151), so the exporter holds no GitHub App key. github_app.py is gone.

Off by default. traceExporter.rerunFailed.dispatch: false. Enabling it needs serge-dashboard-token in Infisical (the same value as serge's DASHBOARD_API_TOKEN). sergeUrl defaults to http://serge.serge.svc.cluster.local.

Rollout: deploy serge#151 first → merge this → caller PR in transformers → deploy the dashboard with dispatch: false → enable → end-to-end check.

🤖 Generated with Claude Code