Rows were read via peek(), which gives Paging no access hints, so
prefix matches beyond the first page never loaded. Drop the head/tail
paging: a query reads the whole prefix range of every active
dictionary in one scan, merges and ranks it, and hands Paging a single
page. The append error row and its strings go away; the prefix scan
reuses search() with substring matches disabled. FileSource reads
through a buffer, cutting per-entry syscall overhead.
OneWay
a4197d2970