PyCForge 1.3.87 / PCFC11 0.2.34 / PyCForge SDK 0.6.4 unchanged
November 5, 2026 — unpublished development handoff

Completed roadmap increment

PCFC11 now supplies bounded reaching storage definitions and bidirectional use
chains over its verified shared integer control-flow graph. The result identifies
which tracked assignments or unknown definitions can reach a read across joins,
loops, switch dispatch/fallthrough and direct labels. Entry storage is explicitly
unknown. Opaque effects conservatively consume incoming definitions and create
unknown clobbers; these are not claims that the original operation writes every
local. The solver publishes only a completed fixed point and complete chains.

Invalid graphs, allocation failure and exhausted size/work limits return no
result and preserve the IR/AST. Existing IR limits remain 64 slots, 512 blocks
and 4,096 instruction entries; this analysis additionally caps definitions at
8,192 and use links at 65,536. It does not diagnose uninitialized reads or supply
alias analysis. Normal compiler transformations do not yet consume these chains;
no speedup, dead-store removal, code motion or wider syntax is introduced.
See the pinned source's docs/DEFINITION_USE.md for the contract and reproduction.

The compiler pin is PCFC11-0.2.34-source.zip (6,018,064 bytes, 843 files).
SHA256: fd3ad9a67f53c2e6644e82c872d9c32eae27906232965328da755786aa525487

README and SDK scope

The main README changes only its version heading to 1.3.87. Its product text,
section categorization, four images and one SDK PyPI link are preserved. SDK
0.6.4 and its README are unchanged; the README still matches its wheel metadata
byte for byte. Crucible and the math/NumPy/JSON contracts retain their existing
semantics. No SDK bump is needed for this compiler-analysis foundation.

Application suite

The complete corrected unfiltered run passed 3,908 cases, skipped 39
and failed zero, accounting for all 3,947 collected cases. It used four
workers, the pinned PCFC11_PATH, Qt offscreen and a 180-second outer watchdog,
and completed in 621.31 seconds. No test was deselected or assertion weakened.
Twenty skips require unavailable Clang, 18 require sealed historical artifacts,
and one requires an environment without PySide6. The last case passed separately
in the SDK environment without Qt. Two QDragEnterEvent deprecation warnings remain.

full-accounting.json derives the result from unique individual JUnit testcase
outcomes and cross-checks the terminal summary. The report's aggregate tests
attribute is 8,943; it is not the number of actual cases. The original XML
is preserved, including that reporter discrepancy.

Compiler qualification

- The independent backward stopping-path oracle covers 2,197 exhaustive ordered
  three-block topologies and 512 seeded graphs: 2,709 graphs and 11,683 expected
  use links. Direct chain symmetry, ownership, immutability, exact size/work
  bounds and every allocation-failure position are checked.
- All seven direct configurations pass 903,603 assertions each: GCC O0/O2,
  PCFC11 O0/O1, GCC ASan/UBSan and two Windows instruction-model optimization
  levels. The AST-to-IR builder suite passes 268 checks in each of five native
  and sanitizer configurations. Models do not qualify native Windows guard pages.
- All 43 standard Linux phases pass. Bootstrap/self-compilation and inherited
  tests pass all 36 recorded steps; the sanitizer build and 33 suites pass all
  34 steps. Eight additional reference/toolchain gates pass.
- The independent Windows-target C assembly preflight passes 326 checks on Linux.
  These are finite correctness checks, not a performance or full-C11 claim.

Fresh applications, memory and installed product

All 53 freshly converted source bundles pass 265 native execution variants and
8 production BuildService profiles. The independent complete-application outcomes
and failed/stale/canceled build-preservation contracts remain qualified within
their declared scope. Thirty-nine PYC4699 Runtime-preview warnings are recorded.
The installed managed compiler produces Windows assembly for all 53 bundles at
O0/O1: 106 checks, with unique assembly files and per-case hashes retained.

The 40 regenerated GC fixtures preserve the predecessor C bytes and pass 80/80
checks with PCFC11 and 80/80 with GCC on Linux. Installed wheel checks compare all
997 product files, build PCFC11 through ManagedCompiler, verify the recorded
identity and execute converted Python plus direct switch-entry programs at O0/O1.
Installed Qt application/startup/About and all Help-footer versions pass.

All eight source gates pass: Ting, promotion, README rendering, SDK catalogue,
bytecode, Ruff, dependencies and conversion documentation. The documentation gate
checks 32 Help pages, 82 Python examples, 50 Runtime and 32 Typed native
comparisons, five API examples, six C callers and three supplied-module examples.

SDK compatibility after the complete application pass

The separately installed, unchanged SDK 0.6.4 completes all seven starters with
each of PCFC11 and GCC through the installed PyCForge wheel. Each driver checks
119 native fixtures and 82 expected rejections, including independent array/value
namespaces. Both workflows check readiness, repeated identical package bytes,
explicit registration and byte-preserving rejection after an author edit.
These automated workflows do not substitute for an unaided human author trial.

Retained failed attempts and corrections

The initial full application run passed 3,907 cases, skipped 39 and failed one
packaging test: MANIFEST.in still explicitly named PCFC11 0.2.33. The manifest
now includes the 0.2.34 archive. The unchanged packaging module passed 11 tests
and 14 subtests; the subsequent complete rerun supplies the clean result above.
No application or compiler behavior change was needed for that packaging fix.

The first builder test helper was named read, conflicting with POSIX read in the
included system header. Renaming only that helper to load_var fixed compilation;
no expected definition, use or assertion changed. The initial solver-only run,
builder compilation failure and final successful results remain in the evidence.
An internal pin was rebuilt to correct compiler instruction-page phase counts.
compiler-package-final.json is authoritative; the earlier descriptor is retained.

The fresh-application subprocess initially saw an old editable distribution
metadata label (1.3.86). Its report explicitly says that distribution was not
used for conversion. The parent report records and checks the actual imported
1.3.87 source tree and fresh generated-C hashes. Subsequent installed checks use
the separate 1.3.87 wheel and match every product byte. This is recorded without
rewriting the original subprocess evidence.

Platform limits and handoff

The new native execution evidence is Linux. Windows instruction models and
assembly generation do not establish native Windows execution. CC's successful
Windows feedback concerns preceding builds and remains separately recorded;
no new native Windows logs were supplied for this increment. Leak detection is
not qualified on this host: ASan/UBSan uses detect_leaks=0. SDK native-verification
scope, unaided author testing and the reported Windows display follow-up remain
as documented. Finite results do not guarantee arbitrary C/Python programs.

Evidence is under evidence/pcfc11_def_use_1_3_87; start with EVIDENCE_INDEX.txt.
The complete development ZIP preserves predecessor paths and historical evidence.
The lean source distribution retains maintained tests, fixtures, tools, licenses,
instructions and only the current compiler pin. Final exact-pair checks, a wheel
rebuild from the source distribution, a fresh source-distribution installation,
archive preservation, sizes and SHA256 values are recorded in the accompanying
delivery report. CC alone decides and performs publication.

Next roadmap work

Make backward liveness reuse the verified shared topology and explicit effect
contracts. Compare it with independent path oracles and the existing separate
liveness implementation before changing store elimination or register allocation.
Measured large-function lowering, owned two-dimensional arrays, further reductions
and native-platform evidence remain in the coordinated queue.
