How do you combine a DynamicFrame collection after relationalise and then select fields for a new dynamic frame?

Using relationationalize to flatten and then join back sounds like a big overkill (also you are assuming the the joins know which columns to use)
I would just flatten using the DataFrame schema, which is that the Flatten visual transform does. e.g.:

def flatten(df, maxLevels=0, separator="."):
    cols_path = []
    # Receives a list with the nested names in order and returns a Spark escaped column selector
    col_path_select = lambda paths: '.'.join([escape_name(path) for path in paths])

    def add_fields(fields, level=1, prefix=()):
        for field in fields:
            # If it's a Struct, and we are supposed to flatten it based on maxLevel config
            if (type(field.dataType) == StructType) and ((level <= maxLevels) or (maxLevels == 0)):
                add_fields(field.dataType.fields, level + 1, prefix + (field.name,))
            else:
                # Add the column path to the list, each element of the path list represents on leve
                cols_path.append(list(prefix) + [field.name])
    add_fields(self.schema)
    # Enforce nested fields to use the full name with the alias, otherwise it will just use the last name
    col_list = [self[col_path_select(cols_path)].alias(separator.join(cols_path)) for cols_path in cols_path]
    return self.select(col_list)