The votingHeader column shared the ParticipationAccount record with the
stateProof blob, the encoded signer context including the Merkle tree of
the state proof keys (2.2 MB for a 3M-round key). SQLite rewrites a row as
one serialized record: with an unchanged length the per-round header update
touches one page, but whenever the header's msgpack length changes (an
integer crossing a width boundary, a field appearing at the first rollover,
about six times per batch) the whole record is re-laid out, freeing and
rewriting ~530 overflow pages of the blob: 1058 WAL frames, then a cover of
the same size by the next erase.

Move the header to a single-row VotingHeader table written with
INSERT OR REPLACE; ParticipationAccount keeps master's columns minus voting.

from the export types script

* erase retired subkeys' page images from the WAL

The row-per-subkey stores delete one voting subkey per round, but
SQLite's write-ahead log keeps the image of every page a transaction
wrote until the log is refilled or truncated.

Add db.Accessor.EraseWAL and run it after every write to both stores.
It checkpoints in RESTART mode, so the next writer starts the log at
frame 1