Windows Rust CI aborted pnpr-auth libsql tests with 0xc0000005 (four
times recently: registration_cap_is_strict_across_backend_instances
three times, sync_server::bulk_revocation_refreshes_a_replica_whose_tokens_went_elsewhere
once).
libsql's LibsqlConnection::drop calls disconnect() and then the inner
local::Connection's own Drop calls disconnect() again. Both see the
drop_ref Arc as unique, so sqlite3_close_v2 runs a second time on a
handle SQLite already freed. Usually it reads freed memory as "closed"
and reports misuse. When another thread reused that memory, it read
garbage and crashed. The tests that crashed run several connections
concurrently, which makes reuse likely.
Upstream fixed it in tursodatabase/libsql#2261 (released in 0.10.0),
which nulls the handle after the first close. pnpr stays on the pnpm
fork because tursodatabase/libsql#2301 is still unreleased, so the fix
is cherry-picked onto the fork's pnpm-0.9.30 branch.
The new test installs an SQLite log callback, drops one connection and
asserts no misuse was logged. It logged two before the fork bump.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
