-- 2.2.0 — MINOR. Graph-kind scan-beam (`ef`) retune + decoupling.
--
-- * The graph kind's scan-time beam width is now its OWN knob,
--   `turbovec.graph_ef` (int, Userset, default 0 = auto = 512), instead
--   of `ef = (candidate_count * 4).max(64)`. Because that old rule was a
--   function of the candidate count, `turbovec.hi_dim_rerank = auto`
--   (whose real job is the flat/IVF exact-rerank WINDOW for dim >= 256)
--   was setting the graph's beam as a side effect. Symptom: an INVERTED
--   recall cliff -- 128d, below the rerank threshold with the beam stuck
--   at the bare 64 floor, was the WORST at R@10 0.720, while 512d looked
--   perfect at 1.00 -- and a recall collapse at EVERY dim with
--   hi_dim_rerank = off. Recall was never dim-dependent, only
--   beam-dependent.
-- * The 512 auto default is set from a measured 1M-scale
--   recall-vs-latency frontier on SIFT-128 and GIST-960 (32 vCPU
--   AVX-512, exact-cosine GT); see docs/GRAPH_EF_BENCH.md.
-- * `hi_dim_rerank` keeps doing its real job for flat/IVF, unchanged.
--
-- Wire format UNCHANGED (stays v8). NO REINDEX: the beam is resolved at
-- scan time, so existing graph indexes get the new default immediately.
-- Flat and IVF indexes are unaffected (no beam on those paths).
--
-- No SQL-surface objects are created or altered by this release: a GUC
-- is registered by the shared library in _PG_init, not by SQL. This file
-- exists to satisfy the per-release migration-coverage invariant
-- (scripts/drift-check.sh, migration_files_cover_documented_versions).
--
-- Upgrade: ALTER EXTENSION pg_turbovec UPDATE TO '2.2.0'; then restart
-- the backend so _PG_init re-runs and registers turbovec.graph_ef.