feat(payments): reverse refunds issued outside the console

Two holes on the refund path, both found by asking what happens when a refund
does not come from our own `/_gm` button.

The merchant cabinet is a second entry point. An operator can refund there, and
such a refund never passes through our API — so the money went back while the
chips stayed credited, silently. Handle `refund.succeeded`: the refund is
re-read from the API (the notification body is no more evidence here than it is
for a payment), bound back to its order through the payment id recorded when
the payment was minted, and reversed through the same engine. It is idempotent
on (provider, refund id), so the event for a refund the console already
recorded reverses nothing twice.

The reversal engine is full-refund-only by design — it revokes exactly what the
pack funded and rejects any other amount — so a partial refund is recorded as
nothing at all and logged loudly for an operator. There is no non-arbitrary way
to decide how many chips a part-refund costs, and guessing would be worse than
asking a human.

Second hole: a refund can still be canceled while pending, and the ledger is
append-only. Recording on any non-empty refund id therefore risked revoking a
customer's chips for money that stayed with us, with no way to take the row
back. The console now records only a `succeeded` refund and tells the operator
to press again otherwise — the idempotency key returns the same refund rather
than paying twice.

Tests: unit (GetRefund, the refund notification envelope, a non-final status
surfaced to the caller); integration (a cabinet refund is reversed once and a
redelivery is a no-op, the event after a console refund changes nothing, a
partial refund records nothing, an unconfirmed refund reverses nothing, a
pending refund records nothing until it settles and then does).

The suite shares one database and the ledger dedupes refunds globally, so the
fake provider now mints a refund id per payment — a constant id made one test's
refund look like another's duplicate.

Decisions D50 (amended) and D52; the notification subscription list in the
deploy docs gains refund.succeeded.

docs/PAYMENTS.md

@@ -303,14 +303,28 @@ an abandoned pending) is surfaced to the user; "payment succeeded" is a hook (em
 message).
 
 **Refunds.** ToS is **non-refundable** — we do not offer refunds to the user. Refunds are
-**admin-triggered** (the E7 console), since no rail pushes an unsolicited refund. On the direct
-rail the console does the whole job in one click (D50): it calls YooKassa's refund API first
-(`POST /v3/refunds`, the order id as the `Idempotence-Key`, carrying the refund receipt of §12) and
-records the reversal only once the money has actually moved — a failed call records **nothing**, so
-the ledger can never claim a refund that did not happen. The recorded refund id is the provider's
-own, which keeps the ledger reconcilable against YooKassa's records. VK refunds are still handled by
-support and Telegram Stars refunds issued with `refundStarPayment`, both recorded by hand afterwards.
-All of them converge on one engine —
+**admin-triggered** (the E7 console). On the direct rail the console does the whole job in one click
+(D50): it calls YooKassa's refund API first (`POST /v3/refunds`, the order id as the
+`Idempotence-Key`, carrying the refund receipt of §12) and records the reversal only once the money
+has actually moved — a failed call records **nothing**, so the ledger can never claim a refund that
+did not happen. A refund is recorded **only in status `succeeded`**: one still `pending` can yet be
+canceled, and the ledger is append-only, so recording early could revoke a customer's chips for money
+that stayed with us. Pressing again is the way out — the idempotency key returns the same refund
+rather than paying twice. The recorded refund id is the provider's own, which keeps the ledger
+reconcilable against YooKassa's records.
+
+**The cabinet is a second entry point (D52).** Unlike the earlier rails, YooKassa lets an operator
+refund from the merchant cabinet, and such a refund never passes through our API — so the money would
+go back while the chips stayed credited. The `refund.succeeded` notification closes that: the refund
+is re-read from the API (the body is no more evidence than a payment notification's), bound back to
+the order through the payment id we recorded, and reversed through the same engine — idempotent on
+`(provider, refund id)`, so the notification for a refund the console already recorded reverses
+nothing twice. The engine is **full-refund-only** by design (it revokes exactly what the pack funded
+and rejects any other amount), so a **partial** refund is recorded as nothing at all and logged
+loudly for an operator: there is no non-arbitrary way to decide how many chips a part-refund costs.
+
+VK refunds are still handled by support and Telegram Stars refunds issued with `refundStarPayment`,
+both recorded by hand afterwards. All of them converge on one engine —
 the `Refund` method (`internal/payments`): it matches the paid order, appends a **refund** ledger
 row (idempotent on `(provider, provider_refund_id)` — the refund id is distinct from the fund's
 payment id, so the two rows coexist under the same partial-unique index), and **best-effort revokes