Harden XML parsing via commons-secure-xml (#766)

* Harden XML parsing via commons-secure-xml

Create XmlStringLookup's document builder and XPath factories through
org.apache.commons:commons-secure-xml. The secure factories enable
FEATURE_SECURE_PROCESSING and install a non-removable entity-resolver
floor on every parser they produce: external DTD and entity lookups are
resolved to empty content instead of being fetched, and internal entity
expansion is bounded, regardless of the JAXP implementation on the
classpath.

Changes:
- Add the commons-secure-xml dependency (1.0.0-SNAPSHOT until its first
  release).
- Route factory creation through SecureDocumentBuilderFactory and
  SecureXPathFactory in XmlStringLookup when the instance's feature map
  enables secure processing. The documented opt-outs keep their
  meaning: feature maps without secure processing, and the standard
  javax.xml.accessExternalDTD system property (which the secure
  factory's resolver floor would otherwise ignore), keep using the
  plain JAXP factories, so external entity resolution can be restored
  where it is wanted.
- Adapt the secure-path tests to the secure contract: a parser may
  either reject a document with an external reference or parse it with
  the reference resolved to empty content; the tests now assert that
  the external content does not leak into the result instead of
  expecting one fixed failure mode.
- Run the CI and CodeQL builds with -Puse-apache-snapshots (inherited
  from the org.apache:apache parent POM) so the commons-secure-xml
  SNAPSHOT resolves; CodeQL's autobuild receives the profile through
  MAVEN_ARGS.

Assisted-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHgnMnGWHQoH2zD2jFdoMT

* Use the Commons Secure XML 1.0.0 release candidate

Bump org.apache.commons:commons-secure-xml from 1.0.0-SNAPSHOT to 1.0.0
and add the temporary staging repository
https://repository.apache.org/content/repositories/orgapachecommons-1962/
after Central, so the vote gets downstream CI results. Drop the
-Puse-apache-snapshots profile from the CI workflows, which the release
version no longer needs. Remove the staging repository once 1.0.0 is
released.

Assisted-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0167e29ScPEdfzJnEFm95imK

* Bump Apache Commons Secure XML from 1.0.0-SNAPSHOT to 1.0.0

* Drop the insecure XmlStringLookup fallback

Commons Secure XML blocks external resource fetching with an entity
resolver rather than with parser features, so the opt-outs the previous
commit documented never worked: neither a feature map without secure
processing nor the javax.xml.accessExternalDTD system property reaches
the resolver floor. Always create the factories through
SecureDocumentBuilderFactory and SecureXPathFactory instead.

Changes:
- Remove the plain-JAXP fallback from XmlStringLookup.lookup(String),
  along with the DEFAULT_XML_FEATURES and DEFAULT_XPATH_FEATURES maps
  that only existed to select it.
- Document the two aspects Commons Secure XML secures separately:
  FEATURE_SECURE_PROCESSING governs processing limits and remains
  settable, while external DTD subsets and entities are blocked outright
  and cannot be re-enabled.
- Remove the references to the "XmlStringLookup.secure" system property,
  which was never introduced.
- Disable the tests that expect an external entity to resolve, and
  tighten the leak assertion now that the parser no longer rejects such
  documents.
- Shorten the changelog entry and state the behavior change.

Assisted-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y7VfssZLkK11bBY2kVyyY4

* Add missing tests

* Fence external XML references in XmlStringLookup

Commons Secure XML blocks external DTD subsets and entities with a
non-removable entity-resolver floor, and lets a caller opt specific
resources back in by installing a resolver of their own. Reuse the Path
fences the lookup already accepts for that: a caller who declared
trusted roots for the document gets its follow-up resources resolved
from those same roots, and from nowhere else.

Changes:
- Add PathFenceResolver, a package-private EntityResolver that resolves
  a system identifier only when it lands inside the fence and refuses
  anything else, including any non-file reference. XmlStringLookup
  installs it only when the fence has roots, so an unfenced lookup
  resolves no external resource at all.
- Parse with the document's own URI, so relative system identifiers
  resolve against the document as XML requires, rather than against the
  process working directory.
- Add PathFence.isEmpty() to tell an empty fence from a real allow.
- Reword the XmlStringLookup and StringLookupFactory Javadoc: secure
  processing covers two separate aspects, the FEATURE_SECURE_PROCESSING
  limits that callers can still turn off and the external references
  that only a fence can opt back in. Drop the references to the
  XmlStringLookup.secure system property, which was never implemented.
- Move the external-reference fixtures to
  src/test/resources/XmlStringLookup, with the referenced files in the
  root and the documents that reference them one directory below, so
  one pair of files covers both the allowed and the refused case
  depending on where the fence points.

Assisted-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QoqFgJRAAHAHhvJTfDRfUt

* Add a changelog entry for the base URI fix

XmlStringLookup used to parse through DocumentBuilder.parse(InputStream),
which supplies no base URI, so the parser expanded a relative system
identifier against the process working directory rather than against the
document. That defect shipped in 1.5 and was reachable through
StringLookupFactory.xmlStringLookup(Map) with a feature map that left
external references enabled, so record it as a fix of its own rather
than folding it into the Commons Secure XML entries.

Assisted-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QoqFgJRAAHAHhvJTfDRfUt

* Cover the malformed system identifier branch of PathFenceResolver

PathFenceResolver.toPath catches URISyntaxException, IllegalArgumentException
and FileSystemNotFoundException, but every refusal the tests exercised went
through the preceding scheme check instead, leaving that catch untested.

- testMalformedSystemIdIsRefused covers the URISyntaxException: an
  unencoded space, the realistic way a system identifier goes malformed.
- testUnconvertibleFileUrlIsRefused covers the IllegalArgumentException
  from Paths.get on a 'file:' URI that parses and passes the scheme check
  but names no path: an opaque one, and one carrying a fragment.

Both assert on the cause, so a change that made the scheme check reject
these earlier would not silently keep the tests green.

Assisted-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QoqFgJRAAHAHhvJTfDRfUt

* Drop the unreachable FileSystemNotFoundException catch

Paths.get(URI) raises FileSystemNotFoundException only for a scheme whose
provider is not installed, and PathFenceResolver.toPath rejects every
scheme but 'file' before it gets there, so that catch type could never
fire.

Assisted-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QoqFgJRAAHAHhvJTfDRfUt

---------

Co-authored-by: Gary Gregory <garydgregory@gmail.com>