fix(ocr): find the snap's bundled tessdata under its tesseract prefix

[nelsonduarte]

Summary

OCR is broken in the published snap (rev 43, 1.15.0): _find_tessdata()
returns None, TESSDATA_PREFIX is never set, and tesseract falls back to
its compiled-in /usr/share/tesseract-ocr/4.00/tessdata, which inside the
snap resolves against the core22 base, where no Tesseract data is installed.
OCR fails with "Error opening data file".

The snap stages Ubuntu 22.04's tesseract-ocr debs, so the data lives at
$SNAP/usr/share/tesseract-ocr/4.00/tessdata. This PR adds a versioned
lookup under the binary's own prefix,
<prefix>/share/tesseract-ocr/<version>/tessdata (newest first, Linux and
macOS only), right after the adjacent-tessdata probe.

• _find_tesseract / _find_tessdata move to app/tools/_ocr_paths.py
  (no Qt or app.* imports of its own); app.tools.ocr keeps both names
  bound.
• tests/test_ocr_tessdata.py: 34 tests, including a pattern-aware fake
  glob and an import allowlist for the new module.

Behaviour change

/usr/bin/tesseract gets the same lookup as before, and Windows is
unchanged. A tesseract under another prefix that carries its own versioned
tessdata (e.g. /usr/local/bin/tesseract with
/usr/local/share/tesseract-ocr/<v>/tessdata) now gets that directory
ahead of /usr/share, even when it holds fewer languages.

Verification

• Published snap rev 43 extracted and run over the core22 rootfs (rev 2955)
  in a container, with the snap's environment and Python 3.10: current
  main returns None and fails; this branch finds
  $SNAP/usr/share/tesseract-ocr/4.00/tessdata, OCRs to .txt and .pdf
  with the right text, and lists deu, eng, fra, osd, por, spa.
• Parity, old vs new _find_tessdata, over 638,976 layout/binary/platform
  combinations: differences only where the binary sits under /usr/local,
  /opt/homebrew or the snap with a versioned tree under its own prefix.
• Mutation testing on Windows, Linux without Tesseract and Linux with
  Tesseract: every behavioural mutant is killed.
• Full suite: Windows 809 passed / 2 skipped; Linux 808 passed / 3 skipped
  (with and without Tesseract installed).
• PyInstaller build: app.tools._ocr_paths is in the bundle.

Not exercised: real snapd confinement (AppArmor/seccomp), a core24 snap,
macOS.

Known issues, not addressed here

• The OCR language combo is rebuilt inside the run and resets to its first
  entry, so with this fix the first OCR in the snap runs in German whatever
  was selected. Pre-existing; separate fix to follow before the release.
• Ghostscript in the snap cannot find gs_init.ps (needs GS_LIB), so
  Compress silently skips its Ghostscript pass. Same class of defect;
  separate fix.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying pdfapps with    Cloudflare Pages

Latest commit:	fde39cb
Status:	✅  Deploy successful!
Preview URL:	https://17212479.pdfapps.pages.dev
Branch Preview URL:	https://fix-snap-ocr-tessdata.pdfapps.pages.dev

View logs