net: end SocketServer's accept flow on a listener that can no longer accept

[Yaraslaut]

Fixes #872.

What was wrong

SocketServer::Core::acceptFlow returned only on Cancelled. Every other failed accept() waited 50 ms and tried again. Some listeners fail every attempt in the same way:

• one that has stopped listening (EINVAL)
• a closed or foreign descriptor (EBADF/ENOTSOCK)
• a non-stream socket (EOPNOTSUPP)

For those, the flow woke every 50 ms forever, and port() kept reporting a port that nothing accepted on.

Reproduced at the current pin. #872 described this as only inferred, arriving with core-cpp ≥ 0.5.1's BadHandle. It already happens with core-cpp 0.5.0, the version pinned today. That version reports every failed accept as SystemError carrying the errno, and the loop retried that too.

The fix

• morph::net::detail::acceptListenerIsDead() covers the failures that recur on every attempt. It accepts both core-cpp spellings: BadHandle/Unsupported, and SystemError with EINVAL/EBADF/ENOTSOCK/EOPNOTSUPP. The supported range (MORPH_CORE_CPP_VERSION 0.5) includes 0.5.0, so both are needed.
• On a dead listener the flow ends and listenerLost() drops the listener. port() then reads 0, which is what the docs already promise when not listening, and a later listen() binds again. Connections already accepted are untouched.
• Exhaustion (EMFILE, ENFILE, ENOBUFS, ENOMEM) is still retried. The existing "survives accept() running out of fds" test still passes.
• docs/spec/core/backend.md now documents which accept failures are retried and which end the flow.

Verification

Measured locally with clang 20.1.2, Debug, the full strict warning set, and core-cpp 0.5.0 (the pin):

• Integration test: "a listener that stops listening ends the accept flow instead of retrying it". It calls shutdown() on the server's listening descriptor underneath it, then checks three things:
  • port() reaches 0
  • a client connected beforehand still gets a reply
  • listen() works again
• Table test for acceptListenerIsDead. It checks each condition three ways:
  • dead: both spellings
  • retried: exhaustion
  • retried: per-peer codes, even when they carry an errno that would mean dead under SystemError
• Mutation 1, the fix reverted (the header from master): the integration test fails, because port() never reads 0 within 5 s.
• Mutation 2, only the BadHandle/Unsupported check with no errno fallback: the integration test also fails. So at the current pin, the fallback is the part that does the work.
• Branch coverage, from clang 20 coverage over morph_net_tests: every arm of the classifier, and of the flow's new branch, is taken both ways. The first push missed include/morph/net's 88% floor on untested arms; that is what the table test and the second commit fix.
• The whole morph_net_tests suite passes: All tests passed (1223 assertions in 216 test cases).
• CI is green on the head, including clang-coverage, clang-tidy-diff, clang-format, the sanitizers and Windows.

Not verified: macOS/BSD. The integration test relies on Linux's EINVAL from accept() on a shut-down listener; morph::net is POSIX-only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WeDnphcJ7EZBLwVfE6pqZu

[codecov]

Codecov Report

❌ Patch coverage is 93.75000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
include/morph/net/socket_server.hpp	93.75%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!