Vector docs: no unbacked INT8 recommendation, document recall at the default beam - #499

Closes 498

Task 1: "INT8 is the recommended setting" had no study behind it (defaults audit S3 says best value NEEDS-AB; the quant-matrix run was void). docs/guide/vectors.md, docs/api/database.md and the docstrings in core.py and vector.py now say that INT8 is what create_vector_index() passes while the engine defaults to no quantization, and that no finished study compares the two. The docstring claim "speeds up search" is removed for the same reason.

Task 2: the "80% vs 99.7% at 20k" figure in Tuning Knobs cited no study and is removed. A new section "Recall at the default beam" reports the measured figures from the 2026-10-24 default-recall study (SIFT recall@10 0.984/0.929 at 10k/100k with the default beam against 1.000/0.998 at efSearch 100; random 1024-d self-hit 295/229/174 of 300 at 10k/50k/100k against 300/297/282) and the footprint study (stored vector first in 295-299 of 300 queries at 10k, 168-182 at 100k). It links ArcadeData#10704 (open; the studies found the cause is the engine's beam policy, not a filter or hop). It is written as an observation, not a recommendation: queries were untimed, only 3 loads per cell, no robustness probes, and on random 1024-d vectors even efSearch 400 reached only 0.40 recall@10 at 100k, so no beam value is recommended.

Checks: pre-commit on the four files, mkdocs build --strict (the Docs Build command). No code block changed, so test_docs_examples.py was not run.

🤖 Generated with Claude Code