[BUG] SocketAddr(u_long, int) performs misaligned sockaddr_in access and truncates out-of-range ports

[thc1006]

Description

SocketTools::SocketAddr(u_long addr, int port) in ext/include/opentelemetry/ext/http/server/socket_tools.h has two defects:

1. Alignment / type-access UB. It binds a sockaddr_in & to the sockaddr m_data storage through a reinterpret_cast:

sockaddr_in &inet4 = reinterpret_cast<sockaddr_in &>(m_data);

sockaddr and sockaddr_in do not share alignment on all platforms (on Linux alignof(sockaddr) == 2 but alignof(sockaddr_in) == 4), so accessing the storage through a sockaddr_in glvalue is an alignment / strict-aliasing issue that UBSan reports.

2. Silent port truncation. port is an int cast to uint16_t with no range check:

inet4.sin_port = htons(static_cast<uint16_t>(port));

so out-of-range values wrap silently:

• -1 → 65535
• 65536 → 0
• 99999 → 34463

Impact

This constructor is on the live server path: it has an addListeningPort caller (unlike the SocketAddr(char const *) overload, which has no in-repo caller). The string-parsing overload was made safe in #4292 (it memcpys a local sockaddr_in in and out of the storage and parses a strict decimal port); this integer constructor still uses the old reinterpret_cast access and the unchecked port cast.

Suggested fix

Mirror #4292's approach: build a local sockaddr_in, memcpy it into m_data, and validate or clamp the port range (or document the caller's contract that the port is already in 0..65535).

Context

Split out of the ext/http correctness audit (#4287). #4292 deliberately scoped this out to keep that PR limited to the string parser.