0.17.6

Fixed

• The File Store no longer reads every subject on every recorded write to check that a record id is unused. A handle builds an index of record ids once — one read of every subject on its first lookup — and keeps it current with its own writes. Measured before the change on 2026-09-03: an import of 517 subjects read 24 GB to write 45 MB. The rule itself is unchanged: a record id reused for different bytes is RecordConflict, identical bytes are an idempotent success, and both hold across handles opened at different times.