[yashaskn8]

Verified defects

1. core/ml/reranker.py rejected SERVY_RERANKER_POOL_SIZE=9/10, making the frozen top-ten pool inaccessible to the V5 reranker even if the operator explicitly evaluated/approved a wider pool.
2. A configured expected checkpoint SHA without a path silently fell back to a different remotely loaded MS MARCO CrossEncoder. That violates pinned-checkpoint expectations.

Repair

• Raise the validated max V5 rerank window from eight to ten, preserving the existing default and current deployment behavior.
• If any expected local checkpoint SHA is provided but model path is omitted, return the original authorized ordering rather than loading a remote legacy model.
• Add synthetic regressions for ten-candidate eligibility, default-five invariance, and missing local pinned checkpoint.
• Document the necessary offline pool 5/8/10 quality/latency tests before changing deployment config.

Boundaries and verification

• No private training data, retraining, model file upload, production feature flag update, or claimed accuracy increase.
• Existing V5 hash preserved.
• Run MySQL backend, frontend build/lint and regressions in CI.
• Production pool increase requires separate evidence of quality and latency.