Reviewed the untracked PESO and LEOPARD indexers and their current happy-path unit tests in apps/vxdata-jobs without editing the monorepo.
Reviewed vxData API and SDK end-to-end test conventions, including the real local PostgreSQL and uvicorn fixture and the query, retrieve, and child_identifiers surfaces.
Designed unit and local integration checks for exact inventory parity, URLs, resource trees, labels, outcomes, conflict detection, and an idempotent second run.
Confirmed exact source and destination counts. Checked the public LEOPARD training label CSV and found 87 recurrence events and 421 censored cases.
Lessons Learned: Pitfalls
A repository-wide text scan initially included generated environments and hid useful results in noise. Restrict searches to source and test trees or exclude .pixi, build output, caches, and worktrees.
The current synthetic LEOPARD fixture uses sequential case identifiers and an artificial 254/254 event split. It cannot pin the released label distribution.
Lessons Learned: Improvements
The vxData ingestion SOP could link directly to the API live_server fixture as the standard local integration setup for SDK-driven jobs.
Dataset indexer tests should record both source inventory composition and destination type counts. A total alone does not localize a missing or extra artifact class.
