PyCForge 1.3.5 — numeric and container conversion extensions
15 October 2026

This release extends the Runtime conversion profile with checked integer bitwise
operators and shifts; integer and floating powers; modular power, including
negative exponents with an inverse; floating floor division and remainder;
divmod; set and frozenset values, operations, methods and comprehensions; list
slice assignment and deletion; dictionary union, unpacking and fromkeys; and
reverse iteration over supported sequences, dictionaries and dictionary views.
List += now accepts any admitted iterable. Mutable container aliases are retained.
Set and frozenset annotations check their current elements at boundaries.

The supported numeric domain remains signed int64 and finite binary64. Overflow,
unhashable elements, invalid slice updates and exhausted resource budgets produce
runtime diagnostics. Set iteration and pop order are unspecified. Dictionary/set
iterator structural-change checks remain stricter than some Python cases. Complex
numbers, arbitrary-size integers, arbitrary protocols and unrelated unsupported
language families are not admitted by these changes. The C-to-Assembly dialog is
not included. Runtime C-IR proof and exact ForgeLens mappings remain unavailable.

The public README is byte-for-byte identical to the approved 1.3.4 README, including
its heading, as requested while conversion upgrades continue. SHA256:
28dec8e51de9f5a3e05ea06ec9528abadcc61f7914e2675a01066ca57ad6fc7f
Package metadata and the application report 1.3.5. HTML Help describes the actual
supported operations in its existing structure. It still has 21 pages and 55
single-file Python examples; no development-status sections were added.

Verification

- All 14 new behavioral test groups pass with GCC AddressSanitizer and
  UndefinedBehaviorSanitizer, including integer boundaries, extreme strides,
  aliases, mixed numeric set keys, modular inverses, mutation, errors and budgets.
  The first sanitizer attempt encountered an environment LeakSanitizer process-
  inspection failure. The successful run disables leak inspection; it does not
  claim a LeakSanitizer pass. Raw logs are retained.
- The documentation verifier passed 55 Python examples (33 Runtime, 22 Typed),
  five API examples, three C callers, two module examples, three README examples,
  Help search/navigation/links, and opening all 21 pages in offscreen Qt. The
  revised container page was visually inspected. The README hash is asserted.
- The first broad unittest run attempted 1,542 tests, with 33 skipped, 28 failure
  records and 14 import errors due to missing pytest in the verification host.
  The newly changed dictionary diagnostic was restored, and seven inventory
  expectations were updated for the three added runtime headers. Rechecking the
  25 affected test methods leaves 20 failure records, with exactly the same
  signatures as the unmodified 1.3.4 baseline. They concern historical version
  and README expectations plus an existing editor module-size budget. They are
  not passing checks and were not hidden or broadly rewritten.
- After installing pytest on the verification host, the 14 previously blocked
  modules passed 190 pytest checks, with 19 skipped. No application dependency
  declaration was changed. The entire broad unittest suite was not repeated
  after these targeted corrections; logs distinguish each run.
- The installed wheel passed both conversion modes, an actual Python/native C
  comparison exercising the additions, and all 21 Help pages. Wheel RECORD,
  embedded README, runtime dependencies, icons and application bytes are checked.
- The source-only corpus contains 16 contributed cases: 4 previously supported
  fixtures remain matches, 5 previously rejected fixtures now match, and 7
  unsupported families remain explicit rejections. Both verified runs have zero
  unexpected outcomes or observed network attempts. Linux seccomp enforcement
  was confirmed using an actual blocked-socket probe. Generated C, source and
  executable hashes and toolchain data are retained. The initial Typed corpus
  attempt treated an existing parentheses warning as fatal; verified runs keep
  the diagnostic and make only that warning nonfatal for Typed. The Typed engine
  was not changed.

The corpus includes schemas, a snapshot of documented conversion tables,
dependency/API manifests, reference sources and immutable run records. These are
fixture comparisons, not universal semantic proofs. No SVM, training dataset,
reusable agents or independent construction-review episodes are fabricated.
The corpus is included in source distributions and excluded from the wheel.

Package size

The new package omits 91 generated screenshots from the 1.1.8, 1.1.9 and 1.1.10
Help-verification directories (18,243,744 bytes before compression). They remain
available in the preceding 1.3.4 package. Their paths and hashes are recorded in
evidence/conversions_1_3_5/omitted-historical-visuals.json. Application icons,
README images, HTML Help, source, tests, licenses and historical text/JSON records
are retained. The source manifest excludes those same screenshot directories so
they are not duplicated inside the included source archive. Build caches and
superseded distribution files are excluded. Current wheel and source archive are
included in dist/. No screenshots from this verification run were added.

Environment and limits

Verified on Linux x86-64, Python 3.12.14 and GCC, with offscreen Qt. Exact local
tool versions are captured in corpus run records. Native Windows/PCFC11 execution
was not performed. Runtime dependencies remain PySide6, numpy, numba, jedi and
autopep8, without added pins. No PyPI publication was performed.

See evidence/conversions_1_3_5 for test and distribution checks, and training_corpus
for the baseline/upgraded comparison. Historical evidence retains its original
version and outcome; this report identifies the current work.
