Deflake JettyTransporterTest.testGet_HTTP3 (UDP port collision)

[elharo]

Fixes #2036

Problem

JettyTransporterTest.testGet_HTTP3 is flaky: binding the HTTP/3 (UDP) connector to the HTTPS TCP connector's port fails with java.net.BindException: Address already in use.

Root cause

The test bound both connectors to the same numeric port by reusing getHttpsPort(), which is an OS-assigned TCP port. TCP and UDP port spaces are allocated independently, so the same-numbered UDP port may already be in use by another process. The collision is inherent to the "let the OS pick a TCP port, then force UDP onto the same number" pattern (verified by experiment: ~1% collision rate with a few hundred UDP sockets occupied).

Fix

• Added HttpServer.findFreeTcpAndUdpPort(), which probes for a port that is free for both TCP and UDP.
• Added HttpServer.addHttp2OnlyConnectorWithMutualTLS(int port) so the HTTPS (HTTP/2) connector can be bound to that exact port.
• testGet_HTTP3 now reserves such a port and binds both connectors to it, keeping the intended "HTTP/2 and HTTP/3 on the same port" semantics.

Verified: JettyTransporterTest (89 tests), ApacheTransporterTest (93), and JdkTransporterTest (92) all pass; the HTTP/3 requests in testGet_HTTP3 are served over HTTP/3.0.

[gnodet]

Clean, well-reasoned fix that correctly identifies the root cause: TCP and UDP port namespaces are independent, so an OS-assigned TCP port may collide with an already-occupied UDP port. The probe-and-retry approach finds a port free in both namespaces.

Observations (non-blocking):

• The TOCTOU race between the probe closing sockets and connectors binding is inherent to any port-probing approach, and the 20-iteration retry loop makes exhaustion statistically negligible (~10⁻⁴⁰ failure probability at 1% collision rate).
• The refactoring of addHttp2Connector to accept an optional port parameter is backward-compatible — all existing no-arg callers pass -1 (OS-assigned).
• Confirmed this is the only place in the codebase that reuses a TCP port number for a UDP binding.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

[github-actions]

@elharo Please assign appropriate label to PR according to the type of change.