Add vectorSearch stage tests

[danielfrankcom]

This change adds tests for the vectorSearch aggregation stage.

The mongot sidecar change here overlaps with #638 and #645. All 3 contain the same commit that should merge cleanly regardless of which is merged first.

I defined a number of pytest fixtures to define the necessary collections for these tests, as an index of this type is far more expensive to set up than the ones we use elsewhere (on the order of 30-60 secs). Using a module/package scoped fixture allows us to amortize this cost across the test cases rather than taking the hit each time.

Closes #376

[documentdb-triage-tool]

🤖 Auto-triaged by documentdb-triage-tool.

Applied: compatibility test, enhancement
Project fields suggested: Component test-coverage · Priority P2 · Effort XL · Status Needs Review
Confidence: 0.30 (deterministic)

Reasoning

component from path globs (test-coverage, test-framework); effort from diff stats (5982+16 LOC, 33 files); LLM failed: Invalid response body while trying to fetch https://api.anthropic.com/v1/messages: Premature close

If a label is wrong, remove it manually and ping @patty-chow so the rules can be tuned. The bot will not re-label items that already have component labels.

[ramgokul-14]

• Question - Pervasive error_code=UNKNOWN_ERROR. 100+ error cases across ~14 files assert the generic sentinel instead of a specific code. This is a compatibility suite; if DocumentDB returns a different code than MongoDB for the same bad input, UNKNOWN_ERROR masks the divergence (REVIEW.md §8, TEST_COVERAGE §6). The PR does use specific codes where it can (VECTOR_SEARCH_LIMIT_NOT_NUMBER_ERROR, EXPRESSION_NOT_OBJECT_ERROR, NOT_FIRST_STAGE_ERROR, etc.), and the comments assert mongot collapses many errors into one opaque code — which may make this defensible. Required action: confirm whether UNKNOWN_ERROR matches a real stable numeric code or means "any error," and add a one-line rationale comment where the executor genuinely can't differentiate. Where a real code exists, capture it in error_codes.py.

• Index-READY polling loop copy-pasted 5–6× (conftest.py, test_smoke_vectorSearch.py, test_stages_position_vectorSearch.py, test_vectorSearch_nested.py, test_vectorSearch_stored_source.py). Extract to a wait_for_search_index_ready(coll) helper in utils/vectorSearch_common.py (REVIEW.md §6). One layer of abstraction is explicitly allowed by TEST_FORMAT.

• Missing documentation updates (REVIEW.md §9). New feature category + new dev-environment dependency: TEST_COVERAGE.md has no rules for search-stage patterns (index polling, mongot dependency, search error surfaces); CONTRIBUTING.md/README.md don't mention the new mongot sidecar that mongo-replset now pulls; FOLDER_STRUCTURE feature tree isn't updated for stages/vectorSearch/.

• searchSyncUser created, never dropped (engine_registry.py). This is infrastructure provisioning (idempotent, once per target — not per-test), so it's defensible, but REVIEW.md §7 calls out user cleanup explicitly

• Fixed dev credential duplicated in three files (dev/compose.yaml, dev/mongot.yml, engine_registry.py). Not a security issue (documented local-dev secret, umask 077), but a drift risk — add cross-reference comments (mongot.yml lacks one).

• nit: test_smoke_vectorSearch.py runs a function-scoped 120s index build for a "smoke" test and lacks a # Property [...] comment; consider reusing the package fixture.

[danielfrankcom]

• Question - Pervasive error_code=UNKNOWN_ERROR. 100+ error cases across ~14 files assert the generic sentinel instead of a specific code. This is a compatibility suite; if DocumentDB returns a different code than MongoDB for the same bad input, UNKNOWN_ERROR masks the divergence (REVIEW.md §8, TEST_COVERAGE §6). The PR does use specific codes where it can (VECTOR_SEARCH_LIMIT_NOT_NUMBER_ERROR, EXPRESSION_NOT_OBJECT_ERROR, NOT_FIRST_STAGE_ERROR, etc.), and the comments assert mongot collapses many errors into one opaque code — which may make this defensible. Required action: confirm whether UNKNOWN_ERROR matches a real stable numeric code or means "any error," and add a one-line rationale comment where the executor genuinely can't differentiate. Where a real code exists, capture it in error_codes.py.

UNKNOWN_ERROR is code 8, which is genuinely correct here. The only thing that differs is the message, and we don't assert messages anywhere in the repo as a deliberate choice, so there's nothing more specific to pin down here.

• Index-READY polling loop copy-pasted 5–6× (conftest.py, test_smoke_vectorSearch.py, test_stages_position_vectorSearch.py, test_vectorSearch_nested.py, test_vectorSearch_stored_source.py). Extract to a wait_for_search_index_ready(coll) helper in utils/vectorSearch_common.py (REVIEW.md §6). One layer of abstraction is explicitly allowed by TEST_FORMAT.

Agreed, extracted this.

• Missing documentation updates (REVIEW.md §9). New feature category + new dev-environment dependency: TEST_COVERAGE.md has no rules for search-stage patterns (index polling, mongot dependency, search error surfaces); CONTRIBUTING.md/README.md don't mention the new mongot sidecar that mongo-replset now pulls; FOLDER_STRUCTURE feature tree isn't updated for stages/vectorSearch/.

vectorSearch is already in feature-tree.csv, and TEST_COVERAGE.md §15 covers stage semantics and error codes generically rather than per stage, so I don't think there's a new rule to add there. The mongot sidecar comes up transparently with the mongo-replset profile and is documented inline in dev/compose.yaml, which README already points to as the single source of truth for targets, so I'd lean toward not duplicating that in CONTRIBUTING/README.

• searchSyncUser created, never dropped (engine_registry.py). This is infrastructure provisioning (idempotent, once per target — not per-test), so it's defensible, but REVIEW.md §7 calls out user cleanup explicitly

That user is provisioned once during replica set initialization rather than per test, and mongot authenticates against it on every run, so dropping it would break replication instead of cleaning anything up.

• Fixed dev credential duplicated in three files (dev/compose.yaml, dev/mongot.yml, engine_registry.py). Not a security issue (documented local-dev secret, umask 077), but a drift risk — add cross-reference comments (mongot.yml lacks one).

The password literal only lives in dev/compose.yaml and engine_registry.py, since mongot.yml reads it from a passwordFile rather than repeating it. There are already comments to cross-reference, and the tests will immediately fail if they don't match so any desync will be caught in the context of review CI.

• nit: test_smoke_vectorSearch.py runs a function-scoped 120s index build for a "smoke" test and lacks a # Property [...] comment; consider reusing the package fixture.

The shared helper above covers the duplicated loop here. Smoke tests don't carry # Property comments anywhere in the repo since those label the parametrized property groups, so the module and function docstrings are the description here. I'd also keep the smoke test building its own small index rather than the package fixture, as it seems clearer for the smoke test to stand by itself.