complete pending uring TCP connects

[russellromney]

Summary

Complete pending uring TCP connects from a shared helper before reporting socket readiness, blocking-connect completion, or SO_ERROR.

This handles the case where a nonblocking TCP connect() is still represented in Quark as TCPConnecting with cached -EINPROGRESS, while the host socket has already completed the connection.

This PR is stacked on #1373 because the regression also needs correct select(2) exception-fdset clearing.

What Changed

• Add UringSocketOperations::CompletePendingConnect().
• Use it from Readiness() for TCPConnecting sockets.
• Use it from GetSockOpt(SO_ERROR) before returning cached connect state.
• Preserve host peer address, make PostConnect() idempotent, and restore TCPInit on completed connect failure.
• Add test/c/soerror_connect_loop.c, a loopback nonblocking-connect regression that checks successful readiness, getpeername(), and refused-connect SO_ERROR clearing.

Proof / Investigation

Diagnostic proof branch: https://github.com/russellromney/Quark/tree/codex/soerror-connect-proof

Tracking issue with the investigation notes: #1374

The proof branch shows the failure without Postgres, psycopg, or application code. It is just a loopback TCP listener and a nonblocking TCP client checking immediate SO_ERROR, select() readiness, and final SO_ERROR.

Validation

Built a qkernel from this branch on the Quark repro host, rebuilt the C regressions as Linux ELF binaries and ran the standalone C loopback regression under Docker --runtime=quark:

summary bad_immediate=0/20 bad_except=0/20 bad_missing_write=0/20 bad_after=0/20 bad_peer=0/20 child_status=0 refused_status=0

Also reran the select regression from #1373 under the same qkernel:

select rc=1 read=0 write=1 except=0 so_error=0

And reran the minimal Postgres 18 + psycopg check:

socket_summary bad_immediate=0/50 bad_select_except=0/50 bad_missing_write=0/50
psycopg_summary failures=0/20