Problem

zeph-memory::document_integration tests fail intermittently in CI with assertion errors unrelated to code changes. Different tests fail on different runs:

• ingested_chunks_have_correct_payload — assert_eq!(all.len(), 1) fails with left: 0 (run #23723176788, PR fix(skills): wire two_stage_matching and confusability_threshold at startup; migrate legacy bundled skills #2410)
• ingest_single_document — similar failure (run #23722894900, main branch)

The failures are non-deterministic and pass on rerun without any code changes.

Root cause

The tests use testcontainers to spin up a Qdrant instance. The race is likely between:

1. Container startup / port readiness check completing
2. Qdrant collection becoming ready to accept writes
3. scroll_all returning 0 results because the ingest point was not yet flushed/indexed

ensure_collection does not wait for the collection to be fully ready before ingesting, and scroll_all may observe an empty state if Qdrant's internal flush hasn't completed.

Impact

• CI flakiness causes spurious PR failures requiring manual reruns
• Blocks CI-gate confidence

Suggested fix

Add a retry loop or readiness probe after ensure_collection before ingesting. Alternatively, use Qdrant's wait=true parameter on upsert to ensure the operation is acknowledged before the scroll query runs.