Reset delivery failures on success even with a stale in-process memo

app/Services/DeliveryHostService.php

@@ -152,6 +152,8 @@ class DeliveryHostService
      */
     public static function recordSuccesses(array $domains): void
     {
+        self::flush();
+
         $flagged = self::flagged();
 
         $domains = array_filter(

tests/Feature/DeliveryHostServiceStaleMemoTest.php

@@ -0,0 +1,60 @@
+<?php
+
+use App\Models\Instance;
+use App\Services\DeliveryHostService;
+use Illuminate\Foundation\Testing\LazilyRefreshDatabase;
+use Illuminate\Support\Facades\Cache;
+
+uses(LazilyRefreshDatabase::class);
+
+/*
+|--------------------------------------------------------------------------
+| DeliveryHostService stale in-process memo - regression
+|--------------------------------------------------------------------------
+|
+| recordSuccesses() filters the success set against flagged(), a per-process
+| static memo (30s TTL). Another worker recording failures for a host cannot
+| invalidate this worker's memo, so a reachable response could be dropped and
+| delivery_failures left un-reset. recordSuccesses() must flush its own memo
+| before reading flagged() so it reloads authoritative state.
+|
+| Simulating a concurrent worker in a single PHP process: we cannot have two
+| separate static memos, so we write the failure rows directly and forget the
+| shared cache key (what the other worker's applyFailure()+flush() does) WITHOUT
+| touching this process's static memo. Calling DeliveryHostService::flush()
+| here would defeat the test by clearing the very memo we are proving stale.
+*/
+
+// Must match DeliveryHostService::CACHE_KEY (private).
+const DELIVERY_HOSTS_CACHE_KEY = 'pf:services:delivery:hosts';
+
+/**
+ * Mimic a different worker recording a failure: bump the DB counter and drop
+ * the shared Redis snapshot, leaving this process's static memo untouched.
+ */
+function otherWorkerRecordsFailure(string $domain): void
+{
+    $instance = Instance::firstOrNew(['domain' => $domain]);
+    $instance->delivery_failures = (int) $instance->delivery_failures + 1;
+    $instance->save();
+
+    Cache::forget(DELIVERY_HOSTS_CACHE_KEY);
+}
+
+it('resets an under-threshold host on success even when the memo is stale', function () {
+    // Prime this worker's memo with the empty (no-failures) snapshot.
+    DeliveryHostService::isUnavailable('a.example');
+
+    // A concurrent worker records 4 failures (still under the default
+    // threshold of 5, so the host is not being skipped).
+    for ($i = 0; $i < 4; $i++) {
+        otherWorkerRecordsFailure('a.example');
+    }
+
+    // The host is under threshold, so delivery to it proceeds normally.
+    expect(DeliveryHostService::isUnavailable('a.example'))->toBeFalse();
+
+    DeliveryHostService::recordSuccess('a.example');
+
+    expect(Instance::whereDomain('a.example')->value('delivery_failures'))->toBe(0);
+});