[kpaxico]

Fixes #1271.

When bind fails in the Server constructor, the error is logged with acceptor_.address() and acceptor_.port(). Both call local_endpoint() without an error code, on the acceptor that the failed bind left unbound. On Windows, the call throws (WSAEINVAL) in the middle of the log statement: the original bind error is lost, startup_failed_ is never set, and app.run() throws instead of returning with is_bound() false. On Linux, it does not throw, but the message reports 0.0.0.0:0 instead of the requested address and port. This PR logs the requested endpoint instead, as the line did before the Unix socket support (45d3fd5). Asio writes both endpoint types to a stream: a TCP endpoint as its address and port, a Unix socket endpoint as its path.

The new test server_port_in_use, next to server_invalid_ip_address, holds a loopback port with a second socket (without SO_REUSEADDR) and serves the app on that port. It captures the log with an ILogHandler and checks that run() returns without an exception, that is_bound() is false, and that the error line reads Failed to bind to 127.0.0.1:<port> - <reason>. It sets the log level it needs and restores the level and the default handler afterwards, so it does not depend on the order of the tests.

Tested

• Windows 11, MSVC 2022 (19.44), standalone Asio 1.24.0: the new test fails before the change (local_endpoint: An invalid argument was supplied.) and passes after it. ctest -C Release passes (139 test cases).
• Linux (Debian 12 container), GCC 12.2, standalone Asio 1.22.1 from Debian's libasio-dev: the new test fails before the change (Failed to bind to 0.0.0.0:0 - Address already in use) and passes after it. ctest passes (crow_test and template_test).

Not tested

• CROW_USE_BOOST and macOS were not tested locally. The CI matrix covers both.
• SSL and compression were off in my local builds. The CI builds with both on.

Thanks!