[yashaskn8]

Scope

• Keep V5 as the preferred frozen local reranker; do not ship private checkpoint weights to GitHub.
• Add an offline, SHA-256-pinned installer for the user-approved V5 best_model directory, refusing symlinks and overwrites and copying into an ignored location.
• Ensure the existing V5 smoke command rejects any other model identity.
• Add six synthetic installer tests and two command regression tests; include installer suite in CI.
• Update local no-Docker Windows deployment steps, distinguish historical 88.10% Hit@1 from latest diagnostic 83.33%.

Verification

• Local synthetic installer unit tests: 6 passed.
• GitHub CI: pending (no private V5 weights are in Git).
• Real V5 inference and end-to-end serving are deliberately not represented as verified; they require the approved checkpoint on the actual serving machine.

Deployment gate

Copy the full V5 best_model from approved Drive storage to the server, run python tools/install_servy_v5_checkpoint.py --source <local-folder>, configure the local .env with digest b1cf23ba0806f0057d73a5ab93fde88d634fc278f71b27f3a2a915afd8f8bcda, run python manage.py verify_v5_reranker, then run the gated real-weight request-path test. No automatic production promotion.