fix(tests): repair PR #12 Windows hook-lifecycle and macOS connection-cap failures

Windows: cli_windows_lifecycle (tests/test_cli.c:9090) installed hooks bound
to <home>/.local/bin/codebase-memory-mcp.exe, but `uninstall` derives its
expected owned-script bytes from cbm_agent_installed_binary_path(), which in
this fork returns codebase-memory-cli.exe. The seeded .cmd scripts and their
extensionless twins therefore never matched the owned shapes and were left
in place. The test now uses the fork's installed binary name, matching the
path uninstall actually resolves. Upstream's test used mcp.exe because its
installed binary still has that name.

macOS: daemon_runtime_connection_cap_covers_slow_hello_and_stopping_is_terminal
keeps a raw connection open without sending hello so it occupies one of two
slots. The fixture's 2 s server hello deadline could expire during the two
full client handshakes on a loaded sanitized runner (the suite took 427 s on
macos-14), freeing the slot before the overflow connect was sampled. This one
fixture now gets a 30 s request timeout through a static override, following
the existing runtime_test_fixture_permanent pattern. The test closes the slow
connection itself, so nothing depends on the shorter deadline. The capacity
assertion is unchanged.

Author: ngallodev