PyCForge 1.3.83 engineering verification
November 4, 2026 · Unpublished development candidate

IDENTITIES AND CHANGE

PyCForge 1.3.83 contains PCFC11 0.2.31 and was exercised with the separately installed, unchanged SDK 0.6.3. Bundled imports remain math 0.3.0, NumPy 0.5.0 and JSON 0.1.0. The compiler archive has 828 files, 5,985,002 bytes and SHA256 7691ebd220577bda3450d4466090ba8e8df02fa8c34c4ea545eca9383305d70f. The SDK wheel has SHA256 15dcca79a7d15483f54034973aa3b3f201935edebfa9fd0be51d46a21a259702.

The public catalogue now includes every admitted bundled operation and its real route decisions. A contract-parity test covers descriptions, guards, lowering, arity and authoring profiles. All 58 bundled exports and two core operations are present. The scalar XML package grammar, workbench schema, package formats and existing extension profiles are preserved. The engine-supplied SDK guide now distinguishes reference browsing, authoring checks, native verification and packaging.

The unchanged SDK's default starter reproduced an actual compiler invocation failure against PyCForge 1.3.82 and PCFC11 0.2.30. PCFC11 0.2.31 accepts the strict flags -fno-fast-math and -ffp-contract=off. Five unsupported options were rejected without changing an existing output. Compiler code generation and optimization were not altered for this change.

Once the driver could compile SDK fixtures, the math starter exposed 34 incorrect signed outputs among 634 output lines in its copysign case. Linux headers selected by PCFC11 expanded NAN as (0.0f / 0.0f). Crucible now constructs a positive quiet binary64 NaN for the single scalar declaration encoding nan, with no floating arithmetic. The focused compiled regression exercises positive sign, unary negation, isnan and FE_INVALID at PCFC11 O0/O1 and GCC O0/O2. The 56 existing math fixtures were left intact.

OPERATOR AND NATIVE RESULTS

Read PyCForge-1.3.83-SDK-Operator-Review.md for the task-by-task human/automation analysis and reproduction commands. Seven starters, tested independently against PCFC11 and GCC, passed 201 declared fixtures per compiler (119 native run cases and 82 expected source refusals). The 14 workflows also check premature and stale builds, package preservation, repeatable bytes and explicit registration. A separately authored metrology package uses an independent binding, namespace, Help and expected outputs. Its combination with records and vectors passed a two-module program on five native configurations, including GCC UBSan. Missing registration produced PYC4702 and no C.

Changing PCFC11's external GCC backend after verification can leave SDK 0.6.3 ready to package. A disposable trial replaced the backend path with a nonexistent one. Readiness and Build succeeded from the old evidence. Fresh Verify rejected native compilation. The connected SDK Help includes the current operating procedure, and ROADMAP.md records the identity check still needed. Controlled authoring checks are distinct from an unaided human usability study.

Standalone PCFC11 passed all 40 Linux feedback phases in the supported ASan configuration with leak detection disabled. The initial default run had three failed sanitizer-dependent CFG phases because LeakSanitizer could not read the process task directory. The driver/core suite also passed through a sanitizer-built compiler and a self-compiled compiler. The inherited stage 1/stage 2 tests ran with O1. These are limited driver and execution checks, not a new claim of complete C11 or Windows coverage.

The maintained whole-program matrix covers 53 fresh bundles (14 Typed and 39 Runtime), 265 native profiles, four complete applications and eight successful BuildService executions. Three failed/stale/canceled build checks per application preserve the existing output. The 39 PYC4699 Runtime warnings are retained and reviewed. Windows targeting is checked separately through 106 assembly outputs on Linux. Native Windows execution, desktop behavior and SDK-host expansion require their own evidence.

PRODUCT, SUITE AND PLATFORM REVIEW

All eight product gates passed, including documentation, live SDK export, phase 16 promotion, Ting, the README, compilation of Python sources, focused Ruff and dependency checks. Documentation exercised 32 offline pages, 82 Python examples, 50 Runtime and 32 Typed native comparisons, five API examples, six C callers and three module examples. The installed application displayed the 1.3.83 title, About badge and 32 Help pages in five categories under Linux Qt offscreen.

The original SDK suite on the current engine and installed predecessors 1.3.60 and 1.3.61 recorded 126/130 headless checks and 26/28 offscreen desktop checks passing. Six assertions use a superseded version, profile or operation count. Separate probes using current expected values pass all six tests without altering the maintained SDK artifact or the raw results.

The unfiltered complete candidate run accounted for all 3,946 independently collected test IDs. It recorded 3,856 passed, 39 skipped and 51 distinct failed cases. Pytest reported 52 failures because one case also had a failing subtest. The run used six workers grouped by file, a 120-second case deadline and persisted test events. Forty-three failed cases already appear in the 1.3.82 full record. In paired isolated trials, seven of the other eight failed on both 1.3.82 and 1.3.83, and the remaining startup recovery case passed on both repeats. The full run remains failed. In a separate attempt a Qt worker aborted while deleting a window with a pending restore. The isolated lifecycle module also aborted on both versions, although that test passed within the completed run. The raw attempts, aborted traces, selected follow-ups and skipped coverage are retained in evidence/sdk_operators_1_3_83. Native Windows and LeakSanitizer remain outstanding. The whole-program and operator passes provide no general application release clearance.

The installable wheel and source distribution contain the current compiler only. The development ZIP retains historical source, previous compiler archives, original failures, corrected runs, qualification tools and operator evidence. Final artifact hashes are recorded in the accompanying release report. CC controls publication.
