benchmarks: correct stale protocol text (lean client claim, EF_SEARCH no-op, int8 build cache, LineItem buckets) #244

The client-path and best-practice audits of 2026-10-07 found four statements that are false at the 26.10.1 tag. lean_http.py says every comparator is driven by a driver with no requests-sized overhead (python-arango is requests, qdrant-client REST is httpx, elasticsearch-py is urllib3). PROTOCOL.md says EF_SEARCH=100 is a no-op for ArcadeDB (an explicit value is a fixed beam of max(k, efSearch) against the adaptive max(2k, 20)) and that INT8 arms get a 100,000 build cache (the INT8 special case was removed; the percent applies to every encoding). FAIRNESS.md says every ArcadeDB type is created with the default bucket count (the document lane's LineItem gets cores - 1). Text only, no row split.