ConnectionManager に probeReplicas() を追加 (Phase 5-1 サブ C-5)

cron からの能動的 health probe を v0.1 でカバーするため、全 read replica
を順次 PDO 接続 + DO 1 ping で検査し、host:port → healthy のマップを返す
public メソッドを追加した。失敗時は connection() と同じ規則で dead cache
に記録する (auth → markPoolDead、それ以外 → markServerDead)。

設計判断:
- dead-cache 状態に関わらず全 replica を probe する。プローブの目的は実時点
  の健全性把握であり、既に dead 化された replica の現在状態も計測対象に
  含める方が運用上有用なため
- healthy probe で既存の dead mark をクリアしない。DeadReplicaCache 抽象は
  mark-alive を持たず、recovery は TTL 自然減衰に委ねる設計
- health_check 設定に関わらず常に DO 1 を発行する。プローブの目的が PDO
  接続確認 + サーバー応答確認の両方を含むため
- probe 用 Connection はインスタンスにキャッシュしない。リクエスト経路で
  使われる replica 選択ロジックを probe 側が汚染しない構造を維持

テスト 14 件追加 (空 read / 全 healthy / 混在 / connect 失敗 / auth 失敗 /
ping 失敗 / dead-cache 透過 + 成功時 mark 保持 / 失敗時 TTL リフレッシュ /
probe 後の selection キャッシュ非汚染 / 既定 pool / 明示 pool 名 /
undefined pool / バリデーション伝播 / port 反映)。

Author: logicraft

src/Sloop/Database/ConnectionManager.php

@@ -32,6 +32,12 @@
  *
  * Empty `read` list collapses replica routing to the primary so single-pool
  * setups keep working without reconfiguration.
+ *
+ * probeReplicas() complements the per-request lazy detection by actively
+ * connecting to every replica and refreshing the dead cache for failures.
+ * Healthy probes never clear existing dead marks (recovery is bound by the
+ * cache TTL), so it is best run from cron to warm the negative cache ahead
+ * of request traffic rather than as a recovery mechanism.
  */
 final class ConnectionManager
 {
@@ -189,6 +195,45 @@ private function getReplicaConnection(string $name): Connection
         );
     }
 
+    /**
+     * Actively connect to every replica in a pool and report each result.
+     *
+     * Each replica is built via the ConnectionFactory and verified with
+     * Connection::ping(); failures at either step are recorded in the dead
+     * cache identically to a request-time miss (auth → markPoolDead, anything
+     * else → markServerDead). Existing dead marks are not cleared on a healthy
+     * probe — recovery still depends on the cache TTL — and the dead-cache
+     * filter is intentionally bypassed so operators can see when a previously
+     * marked replica has recovered. Probe connections are not cached on the
+     * manager: each Connection is destroyed when its iteration ends so that
+     * subsequent connection(writable: false) calls run normal selection.
+     *
+     * @param  string|null            $name Pool to probe; defaults to the configured default pool
+     * @return array<string, bool>          host:port → true (probe succeeded) | false (probe failed)
+     * @throws InvalidConfigException       When the pool is undefined or its config is malformed
+     */
+    public function probeReplicas(?string $name = null): array
+    {
+        $poolName = $name ?? $this->defaultName;
+        $pool     = $this->resolvePool($poolName);
+        $results  = [];
+
+        foreach ($pool->replicas as $replica) {
+            $key = $replica->host . ':' . ($replica->port ?? 0);
+
+            try {
+                $connection = $this->factory->make($replica, $poolName);
+                $connection->ping();
+                $results[$key] = true;
+            } catch (DatabaseException $e) {
+                $this->recordReplicaFailure($pool, $replica, $e);
+                $results[$key] = false;
+            }
+        }
+
+        return $results;
+    }
+
     /**
      * Resolve and validate the pool config for the given name.
      *