>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3614213. --> Reported by: [mably](https://www.drupal.org/user/3375160) Related to !22 >>> <p>The expiry sweep asks a provider what became of a payment before writing it off. Two of the answers it accepts are not true, and both end with a payment in a state nobody actually established.</p> <h3>A reconciliation budget of zero defers every payment for a day</h3> <p>Four places tell an administrator that zero switches reconciliation off and expires on the local deadline alone: the settings form field description, the <code>reconcile_batch</code> comment in config/schema/kessai.schema.yml, the <code>PaymentManager::reconcileBatch()</code> docblock, and docs/architecture.md.</p> <p>What actually happens is the opposite. In <code>PaymentManager::expireDue()</code> a payment carrying a checkout session takes the out-of-budget branch, which calls <code>deferrable()</code> and leaves it pending. With the budget configured to zero that is every such payment, on every run, until each one is a full day past its deadline. Then it is expired anyway and the run logs a warning telling the operator to raise <code>reconcile_batch</code>. So the setting documented as turning reconciliation off instead delays every redirect payment by 24 hours and floods the log.</p> <p>docs/architecture.md contradicts itself in adjacent sentences, which is the clearest sign the intent was never the current behavior: it says a payment the budget does not reach stays pending, and then says to set the budget to 0 to expire on the local deadline alone. With a budget of 0 nothing is ever reached, so the second sentence cannot hold.</p> <p>The code also argues against itself here. <code>PaymentGatewayBase::reconcile()</code> deliberately returns <code>NothingHappened</code> rather than <code>Unavailable</code>, and says why: for a gateway that can never be asked, deferring would hold every payment until the grace period ran out, waiting for an answer that is never coming. A budget configured to zero is exactly that case, and it takes the other path.</p> <p>The engine cannot currently tell the two apart. A budget <em>exhausted</em> during a run should defer, because those payments really will be asked in the next run. A budget <em>configured</em> to zero should not, because they never will be. <code>expireDue()</code> reads a single counter for both.</p> <p>Two existing tests conflate them by using 0 to mean exhausted, so they need rewriting against a budget genuinely spent by an earlier payment: <code>testUnreachedPaymentsAreLeftPending</code> and <code>testBudgetDeferralIsAlsoBounded</code> in tests/src/Kernel/PaymentExpiryTest.php.</p> <h3>A payment the finalizer never examined is reported as NothingHappened</h3> <p><code>WorldlinePaymentFinalizer::finalize()</code> returns immediately unless the payment is of kind <code>payment</code>, which is deliberate: a deposit or a later charge resolves when it is charged, not from a checkout return.</p> <p>But <code>resolveFromProvider()</code> does not know that happened. It calls <code>finalize()</code>, sees the state is still pending, and returns <code>NothingHappened</code>. <code>PaymentGatewayInterface::reconcile()</code> defines that answer as licensing the caller to expire the payment, and says to return it only when the provider actually said so. Here nothing was examined at all, so a deposit or later-charge record that carries a checkout session is written off after a reconciliation that structurally could not have looked at it.</p> <p><code>Unavailable</code> is the honest answer for a payment this finalizer will not examine, and it is also the safe one: it defers rather than expires.</p> <h3>Scope</h3> <p>Both are the same mistake the parent series existed to remove, which is recording an outcome nobody established. Fixing them together keeps the reconciliation contract described in one place.</p> <p>Every claim about what zero means has to move with the behavior: the form description and its French translation, the schema comment, the docblock and docs/architecture.md.</p>