[argszero]

Summary

AppState.db is a single process-wide Arc<Mutex<Connection>>, and four handlers ran argon2 while holding it: login, register, reset_password and change_password. Argon2 is deliberately expensive (~0.24 s with the default parameters), so one unauthenticated request blocked every DB-backed path in the process (including order settlement) for that long.

Measured through the real router with a thread timing st.db.lock():

request	max_lock_wait
login (known e-mail, wrong password — no auth needed)	235.5 ms
register (fresh e-mail — no preconditions at all)	228.2 ms
8 concurrent logins	1.907 s = 8 × ~235 ms, fully serialized
control (unknown e-mail, returns before the KDF)	0.00 ms

A sweep of every non-test db.lock() site found argon2 to be the only slow work running inside a lock, so the fix applies the repo's own rule ("send mail outside the lock", "keep the lock scope to the synchronous section") to the password path.

Related Issue

No linked issue — found while auditing lock hold times. Happy to open one if you'd prefer this tracked.

Changes

• login: read the stored hash under the lock, verify the password outside it.
• register: hash the password (and build the verification code) outside the lock, then re-check the e-mail under the lock before inserting. The e-mail check is repeated because the KDF window now lets a concurrent registration win the race between the first check and the insert; the second check returns the same 409 the first one would have, instead of letting the users.email UNIQUE constraint turn into a 500.
• reset_password: verify the code under one lock scope, hash outside it, and re-check that the code is still valid under the final lock scope before writing the new password.
• change_password: read the old hash under the lock; both argon2 calls (verify the old password, hash the new one) run outside it.
• Removed dao::verify_user_password — find_user_by_email already returns (id, password_hash), so no helper is needed to verify under the lock.
• No configuration or data-structure changes.

Tests

• cargo test — 215 passed, 0 failed
• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• New unit tests added (4)

The new tests measure lock availability with a counter thread that try_locks the real AppState.db once per millisecond while a request runs. The fixture hash carries heavy argon2 parameters (~0.8 s to verify), so the KDF runtime is guaranteed by the PHC string rather than by a timing threshold:

test	observer hits before	after
login_verifies_the_password_outside_the_db_lock	0	158
register_hashes_the_password_outside_the_db_lock	0	159
reset_password_hashes_the_password_outside_the_db_lock	1	525
change_password_hashes_the_password_outside_the_db_lock	0	670

The assertion is a count, not a duration: before the fix each request held the lock from its first query to its last write; after the fix the lock is free during the KDF. A positive control asserts the observer really can acquire the lock while it is idle, so a silently broken observer cannot pass by reporting zero hits. Reverting each handler individually turns exactly its own test red (4 mutations → 4 distinct single-test failures).

Checklist

• Branch naming follows the convention (fix/)
• Commit message uses Conventional Commits (fix(auth): ...)
• Single responsibility, minimal diff (2 files: src/dao.rs, src/routes/mod.rs)