Adapter

• Nothing above the port may name a vendor type — that is the invariant, and it is checkable with a lint rule (What to Automate Out of Review).
• The port is named for what the domain needs, never for the vendor (Naming).
• The fake makes the domain testable in milliseconds, which is the benefit that arrives on day one rather than at the vendor migration.

1// our interface, in our module, in our words
2export interface ChargesCards {
3  charge(amount: Money, method: PaymentMethod): Promise<ChargeOutcome>
4}
5 
6export class StripeAdapter implements ChargesCards {
7  async charge(amount: Money, method: PaymentMethod): Promise<ChargeOutcome> {
8    try {
9      const pi = await this.sdk.paymentIntents.create({
10        amount: amount.minorUnits, currency: amount.currency.toLowerCase(),
11        payment_method: method.token, confirm: true,
12      })
13      return pi.status === 'succeeded'
14        ? { kind: 'captured', id: ChargeId.of(pi.id) }
15        : { kind: 'declined', reason: mapDecline(pi.last_payment_error) }
16    } catch (e) {
17      return { kind: mapError(e), vendorCode: codeOf(e) }   // keep their code
18    }
19  }
20}
21 
22// mapError is a switch over their codes with no default:
23// an unmapped code must fail the build, not the payment.

vendorCode is deliberate: the translation must not destroy the evidence, or every production incident starts with "what did they actually say". Keeping the original code is not a leak — nothing branches on it (Debuggability by Design).