fix(payments): a console refund that its own notification beat is not a repeat

Refunding from /_gm reported "Already refunded" for a refund that had just
succeeded, which reads as "you clicked twice".

The two recording paths race. YooKassa fires refund.succeeded the moment
POST /v3/refunds returns, so the notification handler often writes the reversal
before the console's own write lands. Both name the same refund id, so the
ledger's idempotency index rejects the second — which is the mechanism working
exactly as intended: on the contour the money moved once, one refund row was
written, the balance is right and no abuse flag was raised. Only the message
was wrong about why.

The console now reports success on that path, and the notification's log line
no longer claims the refund was "issued outside the console" when it may well
have come from it.

Covered by an integration test that lands the notification first and then
refunds from the console, asserting the operator is told it succeeded and that
exactly one refund row exists.

.claude/CLAUDE.md

@@ -252,6 +252,11 @@ The native Android app is a Capacitor 8 wrapper of the `ui` SPA, scaffolded unde
     IP bans being prod-only. It bought nothing anyway: the order is resolved from the notification
     metadata **before** any provider call, so a forged id costs one indexed read, and guessing a live
     order id means guessing a uuid.
+  - **The console refund and the `refund.succeeded` notification race, benignly.** YooKassa fires the
+    event the moment `POST /v3/refunds` returns, so the notification often records the reversal before
+    the console's own write lands. Both name the same refund id, so the ledger's idempotency index
+    catches the second and nothing is revoked twice — but the console must not then report "already
+    refunded", which reads as "you clicked twice" for a refund that just succeeded.
   - **Never key a re-check interval off the order TTL.** The order lifetime answers "how long may a
     customer take to pay"; the re-check answers "how soon do we notice a lost callback". Tying them
     together made a failed notification cost the customer the full 30-minute TTL before the chips