[pa-devbot]

Draft handoff for the PaddleOCR crash in C++ CI Serial Programs Mac Intel.

The full investigation lives in MacPaddleOCRCrash_Handoff.md (crash/
control run table, log pattern, what was ruled out, code map, ranked
hypotheses, and concrete next steps including two ready-to-paste CI
snippets for capturing a stack trace).

Summary:

• macos-15-intel segfaults (exit 139) during the first PaddleOCR ONNX
  session creation of the test run, ~0.1s after the
  "Set potential model cache path in session options" stdout line
  (SerialPrograms/Source/ML/Models/ML_ONNXRuntimeHelpers.cpp:28).
  CPU-only sessions (CI default USE_GPU = false); no CoreML involved.
• Probabilistic: 3 crash observations so far (Sockets: add the Qt ServerSocket and the loopback test harness #1787 Mac Intel attempts 1
  and 2; Sockets: add the POSIX ServerSocket #1793 Mac Intel), all with three
  Test_SandwichRecipeDetector tests starting within ~2ms and the socket
  test Sockets (Qt) - Rapid reconnect still running; one run had the
  same overlap and survived. No crash on main, on Mac arm64, or on Sockets: add the WinSocket ServerSocket #1794's
  Mac Intel.
• Ruled out: the socket PRs' code (crash is in the PaddleOCR/ONNX init
  window; socket tests all passed before it; byte-identical socket code
  passed Mac CI earlier), the GPU/CoreML path, model-cache writes (none on
  CPU), and a main regression.
• Also documented while reading the code: a latent cross-language model
  cache race in ML_ONNXRuntimeHelpers.cpp (all PaddleOCR rec models
  share one cache dir + HASH.txt; clean_up_old_model_cache() does
  remove_all on hash mismatch with no synchronization, and
  ensure_paddle_ocr_instance() only serializes per language group).
  Only reachable when the cache is actually written (GPU path / user
  machines), but worth a small standalone fix.

No code changes in this PR - documentation only, for handing off to
another session.

Run IDs and gh commands for every log referenced are in the document.

[pa-devbot]

Closing: the investigation in #1812 found the Mac Intel exit-139 was not in PaddleOCR/ONNX Runtime at all — it was the stack-use-after-return in the Qt server socket's close_connection() completion path, already fixed on main by the shared-ownership change from #1787. This handoff doc is superseded by the findings in #1812.