bugfix: equalise login response time across user-existence branches (0.34.1)

A login attempt against a non-existent username returned 401 in <1ms,
while the wrong-password branch ran argon2 verify (~50-100ms). Timing
the difference let an attacker enumerate valid usernames without ever
seeing a successful response. Run verify_password against a fixed
dummy argon2id hash on the no-user branch so both paths spend the
same compute.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/tests/api_auth.rs

@@ -567,6 +567,91 @@ async fn user_a_cannot_delete_user_b_token(pool: PgPool) {
     assert_eq!(resp.status(), StatusCode::NO_CONTENT);
 }
 
+/// Username enumeration via login response time: an attacker probes
+/// for valid usernames by measuring how long /auth/login takes. Before
+/// the equalisation fix, the no-user branch returned 401 in <1 ms
+/// while the wrong-password branch took ~50-100 ms (the argon2 verify
+/// cost). This test asserts the no-user branch now spends at least
+/// some meaningful fraction of the wrong-password branch's time.
+///
+/// Tolerance is intentionally loose so CI variance doesn't flap the
+/// test. The unequalised gap is large enough (~50x) that even a noisy
+/// CI run with a 5x slack still catches it.
+#[sqlx::test(migrations = "./migrations")]
+async fn login_no_user_branch_runs_argon2_for_timing_equalisation(pool: PgPool) {
+    use std::time::Instant;
+
+    let h = common::harness(pool);
+
+    // Register the victim user so the wrong-password branch has a real
+    // argon2 hash to verify against.
+    let _ = h
+        .app
+        .clone()
+        .oneshot(common::post_json(
+            "/api/v1/auth/register",
+            json!({ "username": "victim", "password": "hunter2hunter2" }),
+        ))
+        .await
+        .unwrap();
+
+    // Warm-up: first login of the process initialises the dummy hash
+    // lazily. Skip that cost when measuring.
+    let _ = h
+        .app
+        .clone()
+        .oneshot(common::post_json(
+            "/api/v1/auth/login",
+            json!({ "username": "victim", "password": "wrong" }),
+        ))
+        .await
+        .unwrap();
+    let _ = h
+        .app
+        .clone()
+        .oneshot(common::post_json(
+            "/api/v1/auth/login",
+            json!({ "username": "ghost", "password": "wrong" }),
+        ))
+        .await
+        .unwrap();
+
+    // Median-of-N is more stable than a single sample.
+    async fn sample_min(
+        app: &axum::Router,
+        username: &str,
+        n: u32,
+    ) -> std::time::Duration {
+        let mut samples = Vec::with_capacity(n as usize);
+        for _ in 0..n {
+            let req = common::post_json(
+                "/api/v1/auth/login",
+                json!({ "username": username, "password": "wrong-guess" }),
+            );
+            let t = Instant::now();
+            let resp = app.clone().oneshot(req).await.unwrap();
+            let d = t.elapsed();
+            assert_eq!(resp.status(), StatusCode::UNAUTHORIZED);
+            samples.push(d);
+        }
+        // Use the minimum: it's the floor that argon2 takes, robust
+        // against unrelated stalls (DB connection acquisition, etc.).
+        *samples.iter().min().unwrap()
+    }
+
+    let wrong_pwd = sample_min(&h.app, "victim", 3).await;
+    let no_user = sample_min(&h.app, "ghost", 3).await;
+
+    // 5x slack: argon2 dominates both branches, so they should be
+    // within an order of magnitude. Unequalised, no_user would be
+    // ~50-100x faster. Asserting "no_user >= wrong_pwd / 5" catches
+    // the bug without being flaky in CI.
+    assert!(
+        no_user * 5 >= wrong_pwd,
+        "login timing leaks user existence: no_user={no_user:?}, wrong_pwd={wrong_pwd:?}"
+    );
+}
+
 #[sqlx::test(migrations = "./migrations")]
 async fn delete_unknown_token_is_404(pool: PgPool) {
     let h = common::harness(pool);