PyCForge 1.3.64 / PyCForge SDK 0.3.0
Documented conversion imports and a shared SDK readiness workflow
Date: 2026-10-28

DELIVERED

The engine publishes ScalarAuthoring.scan and explained documentation capability
metadata. SDK scan and document share the Python service used by the console;
no separate SDK GUI has been implemented. The scan result has a project hash,
engine identity, checklist, generated contract reference, blank author template
and JSON-pointer issues. It never runs examples. Generated blank fields remain
incomplete. Explicit scaffolding preserves author prose and atomically upgrades
legacy projects. Unknown exports, omitted meanings, argument-count mismatches,
placeholder fields, wrong fixture links and oversized Help are rejected.

Every function requires argument explanations, result meaning, details and a
named passing native fixture containing its operation. Constants require their
own explanations; overview, usage, limitations and release notes are required.
The SDK checks documentation before native verification, and verifies linked
example coverage. Documentation edits invalidate previous evidence. Build
requires an exact matching project/engine/SDK/compiler report and preserves
prior artifacts on failure. Repeated builds of one verified snapshot match.

New SDK builds use scalar-package/0.2 with exactly manifest.json, conversion.json
and documentation.json. Both manifest and verification statement bind the Help
hash. The engine checks schema, export coverage, contract identity, budgets and
integrity; it does not run the SDK tests. Admitted packages open through Import
Help in the registration dialog, using the same searchable offline reader.
Only escaped author text is rendered. Examples remain inert. Function links
and example links navigate within the selected package snapshot.

Existing format 0.1 packages remain usable and show a labeled legacy generated
reference. Old engines refuse the new format. SDK 0.3 reads old projects, but
new documented builds require the document upgrade, author completion and a
compatible engine. SDK 0.2 can still browse the final additive engine metadata.
The framework directive 0.3 is shipped in both sources. Engine/scalar semantic
API versions remain unchanged; project/report/package formats evolve separately.

Bundled math 0.2.1 adds complete documentation for the existing 35 functions
and five constants. Its conversion and Help bytes equal the independently
SDK-built math package. The fixed bundled producer label avoids circular
catalogue/evidence fingerprints; serialization itself is not qualification.
Both forms are qualified against the same native fixtures and documentation
requirements. New numeric operations or general array contracts are not claimed.

VERIFICATION

Engine: all 3444 collected test nodes were selected exactly once across four
complete groups. Initial outcomes were 3366 passed, 72 skipped and six affected
nodes failed. Follow-ups resolve all six: two stale release-version assertions,
the new Help-page count, current release-note identities, empty temporary JSON
evidence placeholders, and the IDE module-size guard. Import-specific Help
assembly was moved to its own module; the size guard was not weakened.
Reconciled result: 3372 passed, 72 skipped, zero unresolved failures. This is
complete grouped coverage plus focused follow-ups, not one initially clean run.
The initial run also recorded 4894 passed subtests and one failed subtest; the
module-size failure has a passing follow-up. Subtests are not double-counted
across reruns. Follow-ups: 74 tests/679 subtests, 15 identity tests, 57 catalogue
and distribution tests/516 subtests, and 29 final rendering/admission tests.
The 72 skips retain the previous release's environment/predecessor limits:
sealed predecessor archives, unconfigured PCFC11_PATH, absent Clang, and the
conditional test for an environment without PySide6.

Installed SDK: 60 tests passed in its separate wheel environment. PyCForge and
PySide6 are absent there; the engine environment has no SDK installation. The
final guide/catalogue refresh was then exercised explicitly: old evidence was
rejected without replacing outputs, and the custom and math projects were
scanned, natively reverified and deterministically rebuilt against that exact
final catalogue. The final math verification passed all 48 fixtures (40 native
programs and eight expected source rejections), with complete operation and
linked-example coverage. The independent wave_ops package passed five fixtures,
including a caught TypeError and rejected unsupported calls.

Installed production: the custom package registered, converted and completed
without an SDK dependency. Its C program passed the established warning profile
and UBSan. This profile uses -Wall -Wextra -Werror with the existing exclusions
-Wno-unused-function, -Wno-unused-variable and -Wno-misleading-indentation.
An initial supplemental compile without those exclusions failed on existing
runtime-template warnings; it is recorded and is not claimed as passing.
No runtime C template was changed by this work.

Installed Qt: actual menu/button interaction opened custom and bundled Help,
searched the selected manual, and preserved registration after caller errors.
Unknown members and invalid arity/keywords appeared in both Diagnostics and
the quick toolbar list; activating the list selected source. Dynamic wrong
types retained Runtime exception behavior. Opening the math manual took about
34 ms in the offscreen probe. The screenshot was reviewed internally.

Current Help: 31 pages in five categories; 81 Python examples, 49 Runtime and
32 Typed native comparisons, five API examples, four C callers, three module
examples and three README examples passed. The README retains exactly the four
approved images and matches release 1.3.64.

PERFORMANCE AND PRESERVATION

A synthetic 64-package workload contained 128 functions and 128 constants per
package and about 32.9 MB of documented package bytes. Direct completion on a
short source took 2.04 s on a cold validation cache, then 0.42/0.48 s. With a
480024-character source and a warm cache it took 1.52-1.58 s. Before caching,
repeated short-source requests were about 1.96-2.05 s and large-source requests
3.08-3.15 s. These are direct operation timings on this host, not typing or
render-latency guarantees; they exclude the editor debounce. The existing
process worker and four-second deadline remain in force. The cache stores at
most 128 small content identities, never complete manuals. Current archive,
hash bindings and changed content are still checked; changed Help/contracts
cannot reuse an earlier structural admission.

All 3727 nongenerated baseline source-archive files remain present. Intentional
edits and their paths are recorded in source-preservation.json. Tests, fixtures,
frameworks, useful development materials and historical evidence are retained.
Generated caches/packaging output are excluded from the maintained source and
release payloads. Final distribution, installed-byte and reproducibility
results and exact hashes are recorded in the external handoff report.

SCOPE AND USE

Qualified on Linux x86-64, Python 3.12.14, GCC 13.3.0, Qt offscreen. Windows,
macOS and PCFC11 were not newly qualified. Verification covers the supplied
fixtures and source-call coverage; it is not proof of all inputs, executed
branches, prose accuracy or compiler portability. SDK statements are unsigned.
The scalar route remains positional and Runtime-only; rich parameter schemas,
optional/variadic calls, arrays, NumPy and resource ownership require further
contracts. The separate PySide6 SDK GUI remains future work.

Install the PyCForge and SDK wheels in their respective environments. Existing
registered math overrides persist across upgrades: disable an old override and
enable bundled math, or register the supplied math 0.2.1 package, to select the
new documented version. Use Tools > Register Conversion Imports > Import Help.

The example ZIP includes documented math and wave_ops projects, their sources,
records and packages. Saved evidence is bound to this qualification environment;
run scan and verify locally before building. Qualification scripts retained in
source evidence record the original workspace layout and need paths adjusted
for another host. No release has been published; CC retains publication control.
