refactor(tests): share the OpenAI embeddings mock across vector-store suites

[gold-silver-copper]

refactor(tests): share the OpenAI embeddings mock across vector-store suites

Description

Smell. The vector-store suites under tests/integrations/ (run by test-support/service-tests) each hand-wrote the same httpmock stand-in for OpenAI POST /embeddings. There were 20 copies of one when/then block: the method, the path, the auth header, the request body, and the {"object":"list","data":[{"object":"embedding",...,"index":i}],"model":...,"usage":...} envelope. Around them sat four copies of skip_if_docker_unavailable, two copies of the one-hot embedding helper (axis_embedding and create_embedding_vector), about ten copies of the OpenAI client pointed at the mock, and the flurbo/glarb-glarb/linglingdong definitions, written out in both the mocks and the Word lists. The LanceDB suite also had two identical table-create/IVF-PQ-index blocks.

Why it hurts. A change to the OpenAI embeddings wire meant editing about 20 copies by hand, and the copies had already drifted without anyone choosing it. Some matched Content-Type and some didn't, and the usage numbers varied. In each test, the useful part is which inputs map to which vectors, and that was buried in 30-line JSON literals.

Change. test-support/service-tests/common.rs is a new support module for the service suites, declared once in integrations.rs. It holds:

• mock_embeddings(server, request, embeddings). The expected request body is passed explicitly at every call site, so the request boundary is still asserted, and each test keeps its exact inputs and vectors.
• openai_client(server), skip_if_docker_unavailable, axis_embedding and WORD_DEFINITIONS.

Inside LanceDB, mock_embeddings_api and index_definitions replace the two copies of each block. A FLUMBUZZLE/flumbuzzles() fixture replaces the 256-row padding literal that appeared four times. No assertion changed.

The module lives in test-support/service-tests/ rather than tests/integrations/, because xtask's full_lane_covers_every_integration_suite treats every file in tests/integrations/ as a suite for a rig-<name> crate.

Deliberate small differences

• Every mocked embeddings request now also matches Content-Type: application/json. The LanceDB and Neo4j copies already did, and every suite uses the same OpenAI client.
• Every mocked response reports usage 8/8. Mongo's insert_documents_test used 4/4, and no test reads usage.
• Response index values are now sequential. The OpenAI embeddings decoder ignores index and keeps the array order, so LanceDB's old 0,2,1,1... order decoded the same way.
• LanceDB no longer checks whether the table or index already exists before creating it. Each test uses a fresh TempDir, so those branches never ran.
• Postgres's create_openai_mock_service has #[rustfmt::skip]. Its four recorded-style 1536-value vectors used to sit inside json!, where rustfmt doesn't reach. Outside it, rustfmt would put them one value per line (+6,000 lines).

Net LOC: git diff --shortstat origin/main...HEAD gives 10 files changed, 340 insertions(+), 1003 deletions(-), so net -663.

Rejected or out of scope

• I left the four 7-line WORD_DEFINITIONS → Word { id: docN } maps as they are. Each suite's Word has different derives or serde attributes, so a shared Word type would cost more than it saves.
• The LanceDB chat-completions mock stays inline because it has only one copy.
• Following the scope notes, I left alone the tracing capture helpers (refactor(test-utils): one tracing capture layer for span and event assertions #2659), the recorded_request/recorded_response wrappers and scripted_unary.

Changelog

None (test scaffolding only).

Migration

None.

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation update

Refactor of test scaffolding.

Testing

• cargo nextest run --locked -p rig-service-tests --all-features --run-ignored all with local Docker: 17/18 passed. That covers lancedb (2), mongodb (2), neo4j, postgres, qdrant, sqlite (2), scylladb mock setup and vectorize (6, which skip without credentials). The ignored scylladb::vector_search_test failed at ScyllaDbVectorStore::new with a refused connection to the local Scylla container. That was after its mocked embeddings call had succeeded, so it's a local container problem.
• cargo clippy -p rig-service-tests --tests with all features, no features, and lancedb, sqlite, vectorize and postgres on their own: clean.
• cargo fmt --all, cargo xtask check-test-layout, and cargo nextest run -p xtask full_lane: all pass.

Checklist:

• I have updated READMEs and Rust docs affected by this change (none affected)
• I have added tests that prove my fix is effective or that my feature works (refactor; existing tests unchanged)
• I did not edit CHANGELOG.md or MIGRATING.md (they are generated at release)