-- 2.7.0 — IVF + 1-bit sign-BQ composition; wide-word Hamming kernel;
--         two shipped-in-2.6.0 BQ insert bugs fixed.
--
-- 1. WITH (lists = N, bit_width = 1) now works: cell-contiguous sign codes +
--    coarse centroids + cell directory, so a scan probes only the nearest
--    `turbovec.probes` cells and runs Hamming within them. Wire version stays
--    8 -- the shape is kind = KIND_BQ plus the existing v4 IVF chain fields,
--    so no bump was needed and no existing index is affected. BQ cells live in
--    the RAW L2-normalised space, NOT the rotated space TurboQuant IVF uses:
--    a BQ code is the sign of the centred raw coordinate, so there is nothing
--    to align to, and rotating the query against un-rotated centroids would
--    silently probe the wrong cells.
--
-- 2. TWO CORRUPTION-CLASS BUGS in code shipped by 2.6.0, both fixed here:
--    * `MetaPageData::set_ivf_chains` omitted `bq_mean_count` from its
--      running chain sum, so an IVF+BQ build would have written the
--      coarse-centroid chain ON TOP of the corpus mean -- destroying the
--      centring vector every sign code and every query depends on. This is
--      the FOURTH occurrence of this bug class (v1.24.0 omitted graph_count;
--      2.6.0 fixed three sites). `set_graph_chain` had the same omission
--      (unreachable today, graph+1-bit is rejected) and is fixed too.
--    * Flat-BQ `aminsert` did not re-persist the tombstone bitmap, so every
--      INSERT after a VACUUM silently RESURRECTED every deleted row -- the
--      same M2 bug the graph kind fixed in v2.1.0. It also appended
--      unconditionally, so re-inserting an existing heap TID added a second
--      slot for the same row (unbounded growth under upserts + a duplicate
--      id). Both fixed; the write path now carries tombstones through a
--      single meta write.
--
-- 3. The Hamming kernel counts 8 bytes at a time instead of one (measured
--    4.4-4.8x at embedding dims, up to 5.4x at 1M vectors). No unsafe, no
--    runtime CPU dispatch, so every machine and architecture runs the
--    identical instruction sequence -- the v1.7.3 lesson (a mis-specialised
--    kernel returned WRONG results on pre-AVX2 CPUs). An AVX2 intrinsics
--    kernel was written and PROVEN bit-identical but DECLINED on
--    measurements: only ~1.8x further above dim 512 and a net LOSS below it.
--
-- No wire-format change (stays v8), no REINDEX, no SQL surface change.
-- This migration is intentionally empty.