Flaky JDK 10 CI: ClassicFixtureSmokeTest.convertsIssueXlsxWithCjkText

[shps951023]

Summary

ClassicFixtureSmokeTest.convertsIssueXlsxWithCjkText fails intermittently on the JDK 10 matrix job (test (10, ubuntu-22.04)) in Java CI.

Evidence

• Run https://github.com/mini-software/MiniPdf/actions/runs/35428433188 (merge of PR fix(java): align DOCX rendering with Word for Latin documents #188): JDK 10 job failed once; re-run of the same job passed with no code changes.
• PR run https://github.com/mini-software/MiniPdf/actions/runs/35428425042 (identical head commit): all 18 matrix jobs passed, including JDK 10.
• PR fix(java): align DOCX rendering with Word for Latin documents #188 only touches PoiDocxRenderer.java (DOCX); the failing test is the XLSX path.

Failure details

First assertion fails (ClassicFixtureSmokeTest.java:159), extracted text is mixed CJK + garbage:

限-P--0----0-0期0-0-0修期0--检意重0..0/销￠.0MM0加名称抽N/.检/MO000/MO00材...

Some glyphs extract correctly while others become .notdef-style garbage — consistent with the PdfBoxGraphics2D/EMF text path or font subsetting producing a broken ToUnicode on JDK 10 headless Ubuntu 22.04.

Hypotheses

1. Non-deterministic java.awt font resolution on JDK 10 (fontconfig cache) affecting PdfBoxGraphics2D text drawn from embedded EMF.
2. TrueTypeCollection.processAllFonts picking different faces across runs for multi-face fonts (e.g. wqy-microhei.ttc).

Suggested follow-up

• Reproduce on ubuntu-22.04 + JDK 10 (Zulu) and capture the embedded font + ToUnicode of the failing page.
• Consider pinning the font used by the EMF/Graphics2D path instead of relying on AWT font resolution.