Test isolation + fixed-delay indexing waits make the suite order/load-dependent

[Coriou]

Surfaced by the 2026-06-29 maintenance audit (PR #2). Doesn't bite CI today (fresh server+Redis per run), but makes local re-runs and any future shared-runner setup flaky.

Problem A — registry persistence breaks the rename test on dirty state

reset intentionally preserves namespace registry entries (commit d5ea6d7). As a result, tests/integration/namespaces.test.ts > "rename namespace moves vectors and registry entry" asserts expect(listed).not.toContain("old-ns"), which fails when the server/Redis carries leftover namespaces from a prior run. From a clean Redis the integration suite passes 56/56; against a reused server it fails this one test.

Note: also avoid redis-cli FLUSHALL against a long-running up-vector — its in-memory caches (dimensionMap, knownIndexes) go stale vs the emptied Redis; restart the server after flushing.

Problem B — fixed 500ms indexing waits

A single hardcoded awaitIndexed(500ms) / setTimeout(500) stands in for "RediSearch has finished indexing" across ~20 sites (both setup.ts helpers plus inline in basic.test.ts and several integration files). Under CI load HNSW indexing can exceed 500ms (flaky failure) and wastes time when it's faster.

Suggested direction

• Replace the fixed delay with condition-based polling: awaitIndexed(predicate, {timeoutMs, intervalMs}) that re-issues the query/info() (or checks FT.INFO num_docs) until the expected id/count appears, failing only on real timeout. Centralize in both setup.ts helpers; delete the inline sleeps.
• Make the rename test robust to pre-existing registry entries (assert on the specific moved/old entry rather than its absence from a shared list), or have it reset+verify a unique namespace.

Acceptance criteria

• The integration suite passes regardless of prior server/Redis state.
• No fixed-duration setTimeout/awaitIndexed sleeps remain as the indexing barrier.