PyCForge 1.3.84 — application reliability verification
November 4, 2026 | Linux x86-64, CPython 3.12.14, Qt offscreen
Pinned PCFC11 0.2.31 | Compatibility smoke with separately installed SDK 0.6.3

Application results

The unfiltered application suite collected 3,947 distinct pytest cases. 3,908 passed, 39 skipped and zero failed. Every case had a terminal event. The new test destroys a live Qt owner while its compiler worker is deliberately blocked until cancellation. The previously aborting lifecycle module and the new test also passed together in three separate processes (five cases per run). Pytest's XML includes individual subtests in its test element count. The complete case counts here come from the collection and the per-case event stream.

All 51 failing IDs recorded for 1.3.83 passed together under three-worker execution, under serial execution and in the full suite. PyCForge-1.3.84-Failure-Resolution.txt uses the previous numbering, gives each old observation and explains the repair. The first repair attempt had 49 passes and two failures among those 51. The module-size check led to moving check cancellation into the normal close path. The restore check exposed a test that placed an intentionally short deadline on independent real disk scheduling. The final check blocks the defective read, admits a known-good file and rejects the stale completion.

Application fixes

A passive PCFC11 check can no longer postpone document save, cancellation or exit. Qt owns the queued start. Final owner destruction cancels and joins the worker. An explicit running native build retains its cancel-and-wait close flow. The selected import Help and cached completions now respect pending admission. A document read starts its deadline after validation and any operator prompt.

The import-fault child now loads this source checkout and can exercise real hangs, crashes, malformed responses, cancellation and healthy follow-ups. Import assertions address all three bundled packages and custom registrations by identity. Updated inventory checks retain the exact reviewed assets, actions and converter paths, the prior checkpoint hashes and their tamper tests. A setup dialog may name GCC, while direct tool discovery and subprocess launch remain forbidden in the UI. The supplied Windows GC test kit now verifies its manifest against the project version and the current source fingerprints, with a useful missing-compiler report.

Conversion and native evidence

The pinned compiler was rebuilt from archive SHA-256 7691ebd220577bda3450d4466090ba8e8df02fa8c34c4ea545eca9383305d70f. Crucible freshly converted 53 source bundles. They passed 265 Linux native execution variants across PCFC11 O0/O1, GCC O0/O2 and GCC UBSan. The four complete application examples also passed 8 BuildService profiles. PCFC11 generated Windows-target assembly for 53 cases on this Linux host. The regenerated GC kit passed 80/80 checks with PCFC11 and 80/80 with GCC on Linux. Native Windows execution and the reported Windows display behavior still need firsthand evidence.

All 8 source gates passed, covering release promotion, ting agents, documentation, public README, exported SDK metadata, Python compilation, focused lint and dependency consistency. Documentation verification executed 50 Runtime and 32 Typed native examples. The installed wheel matched 997 product files, rebuilt and checked PCFC11 through the managed build service, and ran the converted example at O0/O1. The installed offscreen window showed version 1.3.84, 5 Help categories and 32 Help pages.

Remaining SDK work

SDK 0.6.3 connected to the installed 1.3.84 wheel, listed seven templates and read all three bundled imports through the public workbench. This was a compatibility smoke. The known SDK package-readiness defect remains open. After successful verification, changing or removing PCFC11’s external GCC backend can leave ready-to-package successful. The next SDK increment should capture and recheck the backend identity, cover changed-path, same-path replacement and missing-backend cases, and then assess human and automated use. The earlier SDK operator review is retained as diagnostic history.

Reproduction

Install the development dependencies in PYPI_RELEASE.md. Set QT_QPA_PLATFORM=offscreen for the Linux desktop tests and supply the built PCFC11 with PCFC11_PATH. Run the complete application suite with python -m pytest -ra. For the recorded parallel run, install pytest-xdist, pytest-timeout and pytest-reportlog, then use -n 6 --dist loadfile --max-worker-restart=20 --timeout 120 --timeout-method thread. Use python -m tools.qualify_whole_program --work-dir <new-directory> --cc gcc --sanitize-oracle for the whole-program matrix. See WINDOWS-GC-TESTING.md for the native Windows handoff.

The complete development ZIP retains the logs and machine-readable records under evidence/app_repair_1_3_84. The companion release report records the final artifact hashes, distribution checks and independent wheel rebuild.
