docs(extraction): report Recall@5 and NDCG@10 on Evaluate on your data

[kheiss-uwzoo]

Summary

• Adds an Accuracy metrics section to Evaluate on your data so readers know NRL end-to-end retrieval benchmarks report both Recall@5 and NDCG@10.
• Keeps Recall@5 as the primary accuracy metric and explains why NDCG@10 is reported as well (public-benchmark comparison; Recall@5 varies more across queries).
• Scoped to this PIC 2026-09-01 decision only. No CLI, eval-harness, or default-pipeline changes.

Why this PR exists

PIC 2026-09-01: Research and NRL E2E will report both metrics going forward. This is not a switch away from Recall. Nave and Even keep Recall as the accuracy north star. Janisha/Sean/Bo asked to include NDCG@10 because popular public benchmarks use it and Recall@5 is noisy.

The BEIR helpers already compute recall and NDCG at cutoffs 1, 3, 5, and 10 (DEFAULT_BEIR_KS). Tests already assert ndcg@10 and recall@5. This PR documents that reporting policy. It does not add a customer eval CLI.

Out of scope (same PIC; separate follow-ups)

• Build endpoint deprecation coordination
• GTC Berlin / AIQ-Retriever GTM after Nave's scope
• LLM NIM SKU split-hosting (Nemotron Ultra / RTX Pro 6000) after Janisha's follow-up
• Nemotron 3 Embed 1B migration / reindex guide
• nemo-memory-datasets partner share
• NIM patch timing for Berlin

Test plan

• python -m mkdocs build --strict --config-file mkdocs.yml from docs/ (local untracked pages moved aside so they did not fail strict)
• git diff --name-only upstream/main...HEAD is only docs/docs/extraction/evaluate-on-your-data.md
• SME glance: Nave/Even okay with "primary accuracy metric" plus dual report in customer docs

Feedback requested

Please focus on the dual-report wording. Confirm we should not name experimental retriever eval / benchmark / skill-eval commands on this page.