handle upstream libxml2 changes (#3056)

**What problem is this PR intended to solve?**

Upstream libxml2 test failures.

- work around a bug/merge-request filed upstream [(on master) fix:
support ASCII encoding on input buffers (!231) · Merge requests · GNOME
/ libxml2 ·
GitLab](https://gitlab.gnome.org/GNOME/libxml2/-/merge_requests/231)
- looks like in-context parser recovery now works? updating
`XML::Node#parse` behavior to accommodate it in a backwards-compatible
fashion. See #2092 which might be fixed in future versions of libxml2.


**Have you included adequate test coverage?**

Making existing tests pass (or not segfault while we wait for an
upstream fix).


**Does this change affect the behavior of either the C or the Java
implementations?**

No.

ext/nokogiri/xml_sax_parser_context.c

@@ -59,6 +59,10 @@ parse_io(VALUE klass, VALUE io, VALUE encoding)
                                (xmlInputReadCallback)noko_io_read,
                                (xmlInputCloseCallback)noko_io_close,
                                (void *)io, enc);
+  if (!ctxt) {
+    rb_raise(rb_eRuntimeError, "failed to create xml sax parser context");
+  }
+
   if (ctxt->sax) {
     xmlFree(ctxt->sax);
     ctxt->sax = NULL;

lib/nokogiri/xml/node.rb

@@ -1049,29 +1049,35 @@ def parse(string_or_io, options = nil)
 
         return Nokogiri::XML::NodeSet.new(document) if contents.empty?
 
-        # libxml2 does not obey the +recover+ option after encountering errors during +in_context+
-        # parsing, and so this horrible hack is here to try to emulate recovery behavior.
-        #
-        # Unfortunately, this means we're no longer parsing "in context" and so namespaces that
-        # would have been inherited from the context node won't be handled correctly. This hack was
-        # written in 2010, and I regret it, because it's silently degrading functionality in a way
-        # that's not easily prevented (or even detected).
-        #
-        # I think preferable behavior would be to either:
-        #
-        # a. add an error noting that we "fell back" and pointing the user to turning off the +recover+ option
-        # b. don't recover, but raise a sensible exception
-        #
-        # For context and background: https://github.com/sparklemotion/nokogiri/issues/313
-        # FIXME bug report: https://github.com/sparklemotion/nokogiri/issues/2092
         error_count = document.errors.length
         node_set = in_context(contents, options.to_i)
-        if node_set.empty? && (document.errors.length > error_count)
-          if options.recover?
+        if document.errors.length > error_count
+          raise document.errors[error_count] unless options.recover?
+
+          if node_set.empty?
+            # libxml2 < 2.13 does not obey the +recover+ option after encountering errors during
+            # +in_context+ parsing, and so this horrible hack is here to try to emulate recovery
+            # behavior.
+            #
+            # (Note that HTML4 fragment parsing seems to have been fixed in abd74186, and XML
+            # fragment parsing is fixed in 1c106edf. Both are in 2.13.)
+            #
+            # Unfortunately, this means we're no longer parsing "in context" and so namespaces that
+            # would have been inherited from the context node won't be handled correctly. This hack
+            # was written in 2010, and I regret it, because it's silently degrading functionality in
+            # a way that's not easily prevented (or even detected).
+            #
+            # I think preferable behavior would be to either:
+            #
+            # a. add an error noting that we "fell back" and pointing the user to turning off the
+            #    +recover+ option
+            # b. don't recover, but raise a sensible exception
+            #
+            # For context and background:
+            # - https://github.com/sparklemotion/nokogiri/issues/313
+            # - https://github.com/sparklemotion/nokogiri/issues/2092
             fragment = document.related_class("DocumentFragment").parse(contents)
             node_set = fragment.children
-          else
-            raise document.errors[error_count]
           end
         end
         node_set