PyCForge 1.3.80 — CFG verification record
2026-11-01, America/Chicago

Scope
PyCForge now pins PCFC11 0.2.28. The compiler's verified integer builder and
backward liveness graph represent short-circuit and conditional truth paths
explicitly in if/for/while/do conditions. Existing data-flow solvers retain
path-specific assignments, and liveness removes only unobserved stores.
Casts, skipped operands, side effects and unchanged resource limits remain
conservative. O0 is unchanged. User labels/gotos, switch, hidden unsupported
control, value-context merges, dominance, SSA, general alias analysis, new
flow warnings and full live-range allocation are outside this increment.

The original PCFC11 0.2.27 archive remains in maintained source and the sdist;
the wheel installs the current pinned source only. All 776 preceding compiler
paths remain in its new source archive. The standalone test-pcfc11.sh workflow,
MIT/upstream attribution, external GCC toolchain and explicit managed build
are retained. The new source has its own version/fingerprint slot. After an
application upgrade, choose Build PCFC11 to build it; existing slots are retained.
No publication, installer, startup build or toolchain download is performed.

Compiler evidence
- Standard Linux feedback: 38/38 phases pass, including every retained suite.
- Bootstrap: 31/31 steps pass, including inherited O0/O1 and 28 suites through
  the O1 self-compiled compiler.
- ASan/UBSan/float-cast-overflow compiler: all 28 suites pass.
- Nine additional reference/tooling gates pass, including sanitized C oracles,
  Windows-target constraints, toolchain paths and harness fault injection.
- The new condition suite executes 96 functions on 64 inputs, comparing 18,432
  result/call-count/trace values per target and optimization level. It adds
  384 independently computed Python path expectations and six O1 assembly
  checks per target. Eighty-one direct builder/solver/liveness assertions pass
  with GCC O0/O2, PCFC11 O0/O1 and a separate direct ASan/UBSan build.
- Unchanged 0.2.27 passes the same execution corpus and fails the explicit-edge
  precision witness. The change improves analysis precision; the witness does
  not identify an old execution miscompilation.

Current application qualification
Fifty-three fresh PyCForge conversions pass 265 native compiler/reference
configurations; all 106 Windows-target assembly checks pass. Four applications
pass eight production BuildService builds/executions and twelve failed, stale
or precanceled output-preservation checks. The current pairwise sum/mean matrix,
complete application, supplied modules and independent C probes pass GCC O0,
GCC UBSan and PCFC11 O0/O1. All 208 scalar results meet their existing exact-
rational error bounds; the documented NumPy rounding differences are unchanged.

189 distinct selected Python cases pass: 134 native build/NumPy cases, 53
source custody/managed compiler/Qt cases, and two SDK workbench inventory cases.
This is a focused maintenance selection, not the full historical repository
suite. The 32-page documentation verifier passes 82 Python examples, five API
examples, six C callers, three module examples and three README examples. The
updated PCFC11 Help page was visually inspected. README heading, five approved
links and banner widths pass. Ting-agent, phase16 promotion, compileall, lint
and environment dependency checks also pass.

Converter implementation and all three bundled import packages are byte-
identical to 1.3.79: math 0.3.0, NumPy 0.5.0 and JSON 0.1.0. SDK 0.6.3 requires
no client update. The engine-owned api.xml changes only its engine version;
index.json updates that version and the corresponding inventory hash. Export
determinism and schema/manual inventory tests pass. No authoring contract changes.

Measurements and qualifications
Five alternating compile samples have mixed sub-0.1-second results. O1 median
current/baseline ratios are 1.23 for parse.c, 0.78 for irconst.c and 1.07 for the
control-flow corpus. Raw elapsed/RSS records remain; shared-host variation and
concurrent reference checks prevent a general speed claim. Extra graph edges
consume unchanged limits and can cause earlier conservative fallback.

Native Windows 0.2.28 execution remains pending. Linux's Windows instruction
model does not qualify PE/UCRT, SEH or guard-page probing. Compiler sanitizers
and sanitized reference programs are separate evidence; detect_leaks=0 was
used, so LeakSanitizer is not qualified. The existing Csmith reference-timeout
skip and target limitations remain recorded. External BLAS/LAPACKE was not
newly qualified. No full C11 or general optimizer-completeness claim is made.

Preserved attempts
Initial direct fixture compilation required a POSIX-name conflict correction
and an allocator declaration. Initial Windows-model linking required its
established model-only stack-probe shim; native outputs already passed. The
SDK inventory check initially detected the expected old engine-version export;
regeneration changed only version/inventory metadata. One workbench test first
failed because the review environment lacked jsonschema; after installing that
test dependency, both unchanged tests passed. No guard or test was weakened.

Evidence is under evidence/pcfc11_cfg_1_3_80, with the compiler's concise records
inside its pinned archive. The original source, historical reports, specifications
and both preceding roadmap snapshots remain. Final wheel/sdist validation,
exact-wheel installed qualification and artifact hashes are recorded separately
in PyCForge-1.3.80-Release-Report.txt after those files are built and frozen.
CC alone publishes releases.
