Dist::Zilla::Plugin::TrialVersionComment fails and times out

[fglock]

Summary

Dist::Zilla::Plugin::TrialVersionComment-0.007 fails under PerlOnJava in
the CPAN compatibility run.

Evidence

• CPAN run: 20260924-085657-82687

• Failing test: t/01-basic.t

• System Perl oracle: all 15 test files passed, 30 tests total.

• PerlOnJava: the test command exceeded the 300-second runtime limit.

• Before the timeout, the plugin produced:

  our $VERSION = '; # TRIAL
  

  The expected output preserves the version literal:

  our $VERSION = '0.001'; # TRIAL
  

The minimal standalone PPI parse/mutate/serialize operation reproduces the
expected output under PerlOnJava, so this appears to require the
Dist::Zilla/PPI integration path rather than being a generic PPI round-trip
failure.

Investigation notes

The repository's PPI support notes already identify unresolved PerlOnJava
object-lifecycle/refcount/DESTROY problems that may explain both the
corrupted transformed content and the long-running test process. This issue
should be reduced to a project-owned minimal reproducer before implementing a
fix; do not change the upstream test.

Acceptance criteria

• Identify the root cause in the Dist::Zilla/PPI integration or PerlOnJava
  runtime.
• Add a permanent focused regression test that passes under system Perl and
  both PerlOnJava backends where applicable.
• Make the plugin's version assignment transformation preserve the original
  version literal and complete within normal test time.
• Re-run the full Dist::Zilla::Plugin::TrialVersionComment test suite.

[fglock]

Proposed fix plan

The first target is the lost version literal in t/01-basic.t. The archived run also exceeded a 300-second limit, but that does not establish an infinite loop or a common cause. Refcount/DESTROY is a hypothesis to test, not an established diagnosis.

1. Establish a reproducible baseline. Record the source commit, built JAR revision, dependency versions, and resolved module paths. Re-run the unchanged upstream suite on system Perl, preserving ordinary optional-test skips (the archived result is 15 files, 30 tests, PASS). Run t/01-basic.t directly with the archived CPAN PERL5LIB overlay on both JVM and interpreter backends, with bounded timeouts and full logs. Record exit status and duration; avoid rebuilding the shared JAR during these runs. If dependencies are missing for system Perl, install them into an isolated temporary library.

2. Reduce the missing integration steps, starting with cloning. Dist::Zilla::Role::PPI::ppi_document_for_file caches documents by the MD5 of encoded content and returns a clone on both cache misses and cache hits. save_ppi_document_to_file serializes the document, updates the file, and caches another clone. The earlier passing PPI reproducer omitted these steps. Compare source and serialized content after parsing, cloning, cache insertion/retrieval, variable discovery, comment insertion, flush_locations, and saving. Exercise both first-use and cache-hit paths. Reduce the first divergent operation to a small reproducer, ideally without Dist::Zilla or PPI dependencies.

3. Identify ownership from the first divergence. Inspect PPI's clone implementation and whichever PerlOnJava primitive it exercises. Check token contents and object identity before/after cloning, aliasing, container stores, and scope exit. Investigate premature DESTROY only if lifetime evidence points there; also check string-copy and value-conversion behavior. Compare bundled versus overlay module resolution. Cross-check PrereqScanner TestNeeds loses quoted module names and version literals #1396, which also loses quoted literals, but link the root causes only if the same reduced behavior explains both.

4. Diagnose the timeout independently. Separate dependency setup, module loading, test execution, and teardown timing. Check host contention and rerun with dependencies already prepared. If a bounded run still stalls, capture JVM thread stacks or a short profile to identify the active code. Establish whether it is slow compilation/loading, blocking, teardown, or non-termination before prescribing a performance fix. Preserve process IDs and clean up only investigation-owned processes.

5. Add regression coverage and implement the smallest justified fix. Add a permanent project-owned test for the reduced defect, validate it with system Perl first, and retain evidence that it fails on the unfixed PerlOnJava parent. Cover preservation of the quoted version, clone independence, and cache reuse if those operations are implicated. Fix the responsible runtime/compiler/provider layer and keep existing upstream tests unchanged. Add a separate bounded regression if a distinct timeout defect is established.

6. Validate and document. Require focused JVM/interpreter success, relevant existing tests for the affected subsystem, and nice -n 19 make before publishing a fix. Re-run the full unchanged TrialVersionComment suite and retain its result, skips, and duration. If PrereqScanner TestNeeds loses quoted module names and version literals #1396 shares the cause, run its reproducer and upstream tests too. Record the proven root cause and remaining limitations in a design document under dev/design/, add a terse Work in progress changelog entry, and run Markdown link checks.

Completion criteria: our $VERSION = '0.001'; # TRIAL is preserved; permanent coverage fails before and passes after the fix on both backends; the upstream suite succeeds with explained platform/optional skips; and the original timeout is either resolved or independently explained with measurements. A larger timeout alone is not evidence of a fix.

[fglock]

Implemented in draft PR #1512.

Root cause: PerlOnJava's Java Clone::clone copied only a scalar's raw fields. That bypassed deferred string-concatenation materialization, so cloned PPI tokens could lose their content. The copier now uses RuntimeScalar's copy constructor for scalar values.

Validation completed:

• system Perl regression test
• JVM and interpreter backend regression tests
• full make
• unchanged Dist-Zilla-Plugin-TrialVersionComment-0.007 suite: 15 files / 30 tests passed, with six expected skips for optional plugins