-- pg_turbovec v1.9.1 -- bench-results-only release.
--
-- This file is a *reference mirror*. The authoritative install
-- script is generated by `cargo pgrx schema`.
--
-- v1.9.1 ships ZERO source-code changes on top of v1.9.0. It bundles
-- the AVX2 latency-frontier benchmark + the honest positioning
-- correction that came out of it:
--
--   * benches/results/latency_frontier_arnold_cohere_1m_v1_9_0_*.json
--     -- the isolated, CPU-pinned, contention-controlled AVX2 run on
--     arnold. Confirms recall@10 = 1.000 on the AVX2 SIMD path and
--     measures the real warm-scan latency (which meh, a pre-AVX2 host,
--     could not).
--
--   * docs/BENCHMARKS.md -- the arnold AVX2 latency section.
--
--   * docs/PARITY_GAPS.md +  -- corrected
--     positioning. The AVX2 benchmark overturned the earlier
--     "we win warm p50" claim (which came from the pre-AVX2
--     scalar-fallback bug fixed in v1.7.3). pg_turbovec is a flat
--     O(n*dim) quantized scan: at 1M x 1024-d its warm p50 is ~2.5 s
--     (AVX2) vs pgvector HNSW's ~5 ms -- HNSW is ~490x faster because
--     it is a sublinear graph traversal. pg_turbovec's durable wins
--     are storage (10-15x smaller), exact recall (1.000 vs HNSW's
--     ~0.96), and build memory -- NOT query latency at scale.
--
-- The wire format does NOT change. The on-disk index format is
-- byte-identical to v1.9.0 / v1.8.x / v1.7.x (MetaPageData::version = 3).
-- There is no DDL or on-disk migration. The upgrade path
-- `ALTER EXTENSION pg_turbovec UPDATE TO '1.9.1';` resolves against
-- this intentionally-empty file. No REINDEX needed.