Two concurrency fixes from the 2026-07-06 review (findings M1, M2):
sync push (M1): the load→merge→store cycle ran as three separate DB
operations. Two devices pushing concurrently both read the same stored
payload, merged independently, and the second store overwrote the first
merge — the server visibly regressed until the losing device pushed
again. The whole cycle (including the leaderboard update) now runs in
one transaction; SQLite serialises the writers. The leaderboard
helper's stale docstring (claiming a single conditional UPDATE that was
actually two statements) is corrected — the transaction now provides
the atomicity it described.
refresh rotation (M2): SELECT-then-DELETE let two concurrent refreshes
with the same token both pass the liveness check and both mint fresh
token pairs. The SELECT is gone; rotation now gates on the DELETE's
rows_affected — whoever removes the jti row wins, everyone else gets
401. Sequential reuse was already covered by
consumed_refresh_token_is_rejected, which still passes unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
funman300