Fix MCP request lifecycle races

changelog.d/unreleased/1736.fixed.md

@@ -0,0 +1,19 @@
+---
+category: fixed
+issues:
+  - 1736
+  - 1737
+  - 1738
+  - 1742
+affected:
+  - src/CodeIndex/Mcp/McpServer.cs
+  - tests/CodeIndex.Tests/McpServerTests.cs
+---
+
+## English
+
+- **MCP server shutdown and request handling now have stronger concurrency guards (#1736, #1737, #1738, #1742)** — request processing now has a bounded timeout response, shared text-writer output is serialized, EOF drains in-flight stdio work with a grace period, and the run-loop flag uses volatile visibility.
+
+## 日本語
+
+- **MCP server の shutdown と request handling の concurrency guard を強化しました (#1736, #1737, #1738, #1742)** — request 処理は bounded timeout response を返し、共有 TextWriter 出力を直列化し、EOF 時は stdio の in-flight 処理を猶予付きで drain し、run-loop flag は volatile visibility を使うようになりました。

src/CodeIndex/Database/DbWriter.cs

@@ -667,7 +667,10 @@ ON CONFLICT(path) DO UPDATE SET
             cmd.Parameters["@checksum"].Value = (object?)file.Checksum ?? DBNull.Value;
             cmd.Parameters["@modified"].Value = file.Modified;
             cmd.Parameters["@generated"].Value = file.Generated ? 1 : 0;
-            return (long)cmd.ExecuteScalar()!;
+            using var reader = cmd.ExecuteReader();
+            if (!reader.Read())
+                throw new InvalidOperationException("SQLite RETURNING id produced no row for file upsert.");
+            return reader.GetInt64(0);
         }
         finally
         {