SCOPE-06..09 cover the completed root id, negation and composition; the cross-space fixture root follows the space-root convention (opaque id = space id).
The suite already regenerates README.md on every run; now a run whose content differs from the committed file also fails, so CI catches a stale matrix. The fresh content is on disk after the run, committing it is the fix.
The refuse specs use registered analyzers (fulltext is gone), the golden regenerates via UPDATE_GOLDEN, MappingGetResp grew an accessor, and the parity suite passes the new NewBackend signature.
The delete-by-query matched the bare Path, so a purge could take same-path documents in other spaces with it; bleve was already RootID-scoped. Pinned as rootscope-04.
The file expansion is grouped again so its NOT stays atomic next to other terms (OpenSearch turned 'mediatype:file OR x' into '(NOT dir) AND x'), and the bleve compiler no longer re-keys resolved groups (a negated mediatype group targeted the raw 'mediatype' field and matched everything). Pinned as MEDIATYPE-07..10.
The engine suites keep what is engine-specific (index setup, health, purge-space batching); every behavior answer lives in the parity suite once. The shared test client learns IndicesCount, and FIELDS-15 pins the Size/Type gate for number queries.
Everything the bleve and OpenSearch backend suites checked now runs against
both engines in services/search/pkg/parity. The two cases without a twin
there join it (a facet value keeps its case, batches stay apart), the one
thing only OpenSearch can do, refuse an unhealthy cluster, stays in its
package.
The expectations are the desired state; where bleve or OpenSearch answer
differently on main today the case records that answer as an engine
override, so the suite is green and the README lists every divergence as
known.
Every case is one spec per engine, so one engine failing leaves the other's
answer in the matrix. Query groups share their engines through an Ordered
container, operations build their own. The README is assembled from report
entries in ReportAfterSuite, which also works across ginkgo -p processes, and
is only written once every planned row got an answer from every engine.
A case can carry, per engine, what that engine answers today where it still
differs from the expectation (engineOverrides): the spec asserts the override,
the README marks the row as known, and the override fails once the engine
answers as expected.
Move the OpenSearch test helpers to services/search/internal/opensearchtest so
the parity package can import them, and drop the index generation and clock
patch the suite was written against, neither exists on main yet.
Dominik Schmidt