Summary

Fixes #3928.

The xml processor and parse_xml bloblang method drop namespace prefixes on element and attribute names during XML→JSON conversion (<dc:title> becomes the key title), which makes the original XML impossible to reconstruct from the JSON output. The root cause is that github.com/clbanning/mxj always uses xml.Name.Local, so there is no flag on the existing call path that preserves prefixes.

This PR adds an opt-in preserve_namespaces boolean to both surfaces (processor field + bloblang param, default false). When enabled, parsing routes through a small stdlib-only walker in internal/impl/xml/package.go that:

• Maintains a URI→prefix scope as it descends, seeded from xmlns:* declarations on each element.
• Emits element and attribute keys as prefix:local when the namespace URI is known, falling back to the raw prefix (which Go's xml.Decoder leaves in Name.Space when a prefix is used but never declared).
• Keeps xmlns:* declarations as -xmlns:prefix attributes so the original XML is reconstructable.
• Preserves the existing output shape otherwise (-attr for attributes, #text for mixed content, arrays for repeated elements), and mirrors mxj's cast order (int → uint → float → bool) for the cast option.

Default behaviour is unchanged, so this is non-breaking for existing users.

Why not wait for upstream Go

The issue links Go CL 355353, which would add namespace-prefix retention to encoding/xml. That CL has been stalled for ~4 years. We don't actually need it: xml.Decoder already exposes the resolved URI in Name.Space and reports xmlns:* declarations as attributes, so a ~70 LOC walker can reconstruct prefixes today.

PoC results — validation on the issue's exact input

Before implementing, I validated three paths on the issue's XML:

<root xmlns:dc="http://my.namespace/dc" xmlns:ot="http://my.namespace/ot">
  <dc:title>This is a title</dc:title>
  <dc:description tone="boring">This is a description</dc:description>
  <ot:elements id="1">foo1</ot:elements>
  <ot:elements id="2">foo2</ot:elements>
  <ot:elements>foo3</ot:elements>
</root>

Path 1 — current mxj.NewMapXml (lossy, reproduces bug):

{
  "root": {
    "-dc": "http://my.namespace/dc",
    "-ot": "http://my.namespace/ot",
    "description": { "#text": "This is a description", "-tone": "boring" },
    "elements": [
      { "#text": "foo1", "-id": "1" },
      { "#text": "foo2", "-id": "2" },
      "foo3"
    ],
    "title": "This is a title"
  }
}

Element prefixes dropped; the xmlns: on the declarations is also lost (attributes appear as -dc / -ot rather than -xmlns:dc / -xmlns:ot).

Path 2 — mxj.NewMapXmlSeq (2-LOC change, but heavy output shift):

{
  "root": {
    "#attr": {
      "xmlns:dc": { "#seq": 0, "#text": "http://my.namespace/dc" },
      "xmlns:ot": { "#seq": 1, "#text": "http://my.namespace/ot" }
    },
    "dc:description": { "#attr": { "tone": { "#seq": 0, "#text": "boring" } }, "#seq": 1, "#text": "This is a description" },
    "dc:title": { "#seq": 0, "#text": "This is a title" },
    "ot:elements": [
      { "#attr": { "id": { "#seq": 0, "#text": "1" } }, "#seq": 2, "#text": "foo1" },
      { "#attr": { "id": { "#seq": 0, "#text": "2" } }, "#seq": 3, "#text": "foo2" },
      { "#seq": 4, "#text": "foo3" }
    ]
  }
}

Preserves prefixes but wraps every element in #attr / #seq metadata — would require every existing consumer to adapt to a new output shape.

Path 3 — custom stdlib walker (this PR):

{
  "root": {
    "-xmlns:dc": "http://my.namespace/dc",
    "-xmlns:ot": "http://my.namespace/ot",
    "dc:description": { "#text": "This is a description", "-tone": "boring" },
    "dc:title": "This is a title",
    "ot:elements": [
      { "#text": "foo1", "-id": "1" },
      { "#text": "foo2", "-id": "2" },
      "foo3"
    ]
  }
}

Prefixes preserved, output shape identical to the current default except for the added dc: / ot: / -xmlns:* keys. Round-trips cleanly to XML.

Tests

Added to internal/impl/xml/processor_test.go:

• TestXMLPreserveNamespaces — 5 sub-cases:
  • Exact XML from The XML processor omits namespaces #3928.
  • preserve_namespaces: true on XML with no namespaces is a no-op vs. default output.
  • Nested element redeclares a prefix (xmlns:a at two levels mapping to different URIs).
  • Namespaced attribute (xsi:type).
  • Prefix used without an xmlns declaration stays literal.
• TestXMLPreserveNamespacesWithCast — confirms cast: true applies to element and attribute values under the new path.
• TestXMLDefaultStripsNamespacesUnchanged — regression guard: omitting the flag keeps the current lossy behaviour byte-for-byte.

Added to internal/impl/xml/bloblang_test.go:

• parse_xml(preserve_namespaces: true) round-trip.
• parse_xml() without the flag still strips prefixes (opt-in guard).

Full test run:

ok  github.com/redpanda-data/connect/v4/internal/impl/xml  0.302s

All existing tests (TestXMLCases, TestXMLWithCast, TestParseXML) continue to pass unchanged.

Test plan

• go test ./internal/impl/xml/ passes
• go vet ./internal/impl/xml/ clean
• go build ./internal/impl/xml/ clean
• Default (preserve_namespaces omitted or false) output byte-identical to previous behaviour
• Issue's XML round-trips with preserve_namespaces: true
• CI green on the PR