PyCForge 1.3.78 — definition navigation and coordinated reliability increment
2026-10-31

Scope and baseline
Built from the complete maintained 1.3.77 source with the coordinated roadmap.
CC reported that Classes & Functions failed whenever an editor contained *import.
The preceding source reproduction raises a syntax error on the import line and
returns no outline. The new release uses Crucible's existing token-aware,
position-preserving declaration preparation in the symbol navigator, source
outline and synchronous function-fold discovery. This reads source as data;
it does not load imports, consult package readiness, or execute Python code.

Declared module imports, aliases, nested definitions, methods, decorators,
Unicode/UTF-16 navigation, CR/LF/CRLF input, original copy ranges and unsaved text
are retained. Valid unknown/disabled namespaces do not block browsing. Genuine
syntax errors and malformed/out-of-preamble declarations remain explicit.
Source size/symbol/fold limits, cancellation and stale-result handling remain.
Both gutter and tab-context entry points are exercised in both editor panes.

Only seven existing package files differ from the maintained baseline: the
three definition observers, workspace Help, product version, and the two SDK
reference identity/inventory files. The other 996 source package files are
identical, including all conversion/runtime code, bundled import archives and
PCFC11 0.2.27 source. The wheel still contains 997 product payload files.
Math 0.3.0 (36 operations), NumPy 0.4.0 (18) and JSON 0.1.0 (two) are unchanged.
SDK 0.6.3 needs no release; engine/catalogue contracts remain compatible.

Source regression evidence
The final selected ledger contains 287 distinct passing source test cases,
with no remaining failure or skip. This is a focused cross-component selection,
not a claim that the complete historical repository suite was run.
- 78 navigator, declaration, outline and workspace cases; two subtests.
- 155 conversion-import, memory, compiler integration, reference and source-
  loading cases. Five sanitizer cases were rerun after the host limitation below.
- 54 folding, editor lifecycle/responsiveness and UI continuity cases; 33 subtests.
Exact case identities and final outcomes are in evidence/navigation_1_3_78/
test-ledger.json; original and final XML/logs are retained.

Documentation and installed application
The final Help verification passes all 32 pages and 82 Python examples (50 Runtime,
32 Typed native comparisons), five API examples, six C callers, three module
examples and three README examples. README edits are limited to the matching
1.3.78 heading and the confirmed definition-navigation bug-tracker row. Rendering
retains exactly five approved URLs, the two banner widths and existing screenshots.

An isolated installed-wheel environment exercises the real Qt editing worker,
source outline and dialog in both panes. The four expected class/function entries
appear; copying preserves text/revision/undo state, and jumping retains correct
UTF-16 positions while revealing the destination. All 997 product payload files
match the wheel and 299 loaded product module origins resolve to that installation.
Offscreen captures are retained and visually reviewed. The installed mixed-import
class workflow prints the expected values under GCC UBSan and PCFC11 O0/O1.
The external release report records the final archive identities and final
installed-wheel/distribution results; release artifacts contain no startup build
or new toolchain installer.

SDK compatibility
The unchanged installed SDK 0.6.3, in a separate environment without PyCForge,
connects to installed engine 1.3.78, reads all default-import forms, and discovers
the existing eighteen array operations. An independent vector namespace passes
all 50 template cases, deterministic packaging and explicit registration. Stale
evidence refuses a rebuild and preserves the last package. The cached workbench
remains available when the engine executable is absent. See sdk-results.json.

Pipeline baseline
tools/benchmark_conversion_pipeline.py retains three reviewed workloads and
records source/C size, conversion time, compilation time, execution, binary size
and available peak-RSS observations separately. The tool bounds complete POSIX
process groups and leaves timed-out memory measurements explicitly unavailable.
It is an explicit development command, not part of editor analysis.

The retained 68,135-byte literal matrix generates 2,016,111 bytes of C in about
2.15 seconds. On this Linux host PCFC11 O0/O1 compile it in about 2.9 seconds and
both executions match the Python/NumPy reference. GCC O0/O1 exceed the 30-second
observation deadline. The compact loop matrix and the practical math/NumPy/JSON
class workflow compile and run correctly under all four profiles. Ten completed
native profiles match their references; two large-case GCC profiles time out.

These are single diagnostic observations per profile, without a speedup claim,
timing acceptance threshold or general compiler ranking. Giant-function code
growth remains open; no optimizer/kernel change is made in this increment.
pipeline-baseline.json identifies its initial and corrected source records.
The roadmap now points next development to the bounded sum/mean contract, with
large-function lowering and native platform follow-up retained as open work.

Corrected verification attempts and limits
The first practical workflow omitted its required declared Runtime field and
correctly raised AttributeError on all compilers. The fixture now declares
values: object; its corrected four profiles pass. No runtime code was changed
to accommodate the fixture. A local name collision in the measurement harness
was corrected before the selective rerun. The first folding-suite collection
found missing Numba/Jedi development dependencies; the complete selected folding
suite passes after installing the ordinary declared application dependencies.

The first final-distribution check correctly rejected a build log that had
continued growing after its sdist snapshot. Final packaging uses a frozen source
staging tree and writes all active logs outside that tree. The validator remains
unchanged; completed evidence is retained.

LeakSanitizer cannot inspect /proc task state in this host. The affected five
cases pass with ASAN_OPTIONS=detect_leaks=0; AddressSanitizer and UBSan remain
active, and the native allocation counters/rollback/active-GC checks remain.
Do not describe this as a successful LeakSanitizer run.

No native Windows desktop execution was available. Offscreen Qt/Linux tests do
not close the Windows compiler, GC or prior Preparing-dialog display follow-up.
On Windows, open a document containing valid *import declarations and classes/
functions; verify both entry menus, Copy, Go To, folds and the split editor.
Current native/GC evidence retains its actual platform and release context.
Only CC publishes. Nothing was published by this task.
