PyCForge 1.3.79 verification record
2026-10-31, America/Chicago

Scope and release identities
This increment implements the coordinated roadmap's next bounded numerical
capability. Crucible and bundled NumPy advance together, with independent SDK
authoring and complete-program/native compiler qualification. PyCForge is
1.3.79; bundled NumPy is 0.5.0 (20 operations). SDK 0.6.3, PCFC11 0.2.27,
math 0.3.0 and JSON 0.1.0 are unchanged. No compiler optimizer or GC changes
are included. CC alone publishes; this work does not publish any package.

Supported behavior
The Runtime array service adds sum(a, axis=None) and mean(a, axis=None) for
existing finite contiguous one-dimensional float64 arrays. Positional and
named arguments are supported; axis is None or an integer 0/-1. The result
is a scalar, does not retain or mutate its input, and requires no helper heap
allocation. This is a shared array capability, also used by independently
named SDK exports. Older profile identities and operation meanings remain.

Accumulation follows a fixed balanced pairwise tree, splitting at floor(n/2),
evaluating left before right and rounding each addition to binary64. Mean
divides that sum by the exact integer count (at most 2**53). Round-to-nearest,
ties-to-even is required; an unsupported rounding mode raises RuntimeError.
Gradual underflow is a prerequisite; fast-math remains refused. A zero total
is canonicalized to +0.0; mean division can produce signed underflow zero.

Empty sum is +0.0; empty mean raises ValueError. Any nonfinite intermediate
sum raises OverflowError, including when later cancellation or division would
give a finite mathematical answer. For nonempty input the helper charges
2*n-1 Runtime steps; exhaustion follows the existing fatal resource policy.
The helper uses logarithmic native stack depth and no internal GC safe point.

No sequence coercion, array .sum()/.mean() methods, dtype/out/keepdims/initial/
where options, views, broadcasting or multidimensional arrays are added.
Unsupported forms remain explicit. This is not general NumPy compatibility.

Numerical evidence and limits
The fixed-tree result is not promised to match NumPy bitwise. Strong
cancellation may produce large relative error: [1e16,-1e16,1] sums to 0 here,
where the exact total and this host's NumPy result are 1. Neither compiler
agreement nor a relative-tolerance comparison alone establishes correctness.
The specification records absolute-error bounds from the tree depth, sum of
absolute inputs, binary64 roundoff and gradual-underflow allowance.

tools/qualify_sum_mean.py exercises 104 deterministic arrays (16,887 values),
including mixed magnitudes, cancellation, odd lengths, subnormals and signed
zeros. Every native result satisfies both the fixed-tree Python oracle and
independent exact-rational bounds. All four compiler profiles agree on all
208 sum/mean outputs. Exactly 52 outputs differ bitwise from NumPy 2.3.5.
Those differences are retained evidence, not silently widened tolerances.

Profiles: GCC 13.3.0 O0; GCC O1 with UBSan, no fast-math and contraction off;
PCFC11 0.2.27 O0 and O1, x86_64 Linux. Each also compiles and executes a
complete class/function/JSON statistics program, a supplied-module program
and an independent C11 floating-point/struct-return/recursive-pointer probe.
All 16 matrix/application/probe executions pass. Compiler identities and
generated-source hashes are in evidence/sum_mean_1_3_79/matrix-results.json.

Selected source and default-import qualification
39 focused cases and 232 selected regressions pass: 271 distinct cases, with
no remaining failure or skip in that selection. This is not a run of every
historical test in the repository. Coverage includes binding, refusal without
partial generated C, catchable errors, input preservation, profile identity,
active roots, memory cleanup, source navigation, completion and offline Help.
The exact cases and corrected final status are in test-ledger.json and JUnit
records under evidence/sum_mean_1_3_79.

NumPy template qualification passes 62 cases: 33 executable and 29 rejection
fixtures. Each executable fixture passes GCC UBSan and PCFC11 O0/O1, for 99
native executions. The NumPy bundle is written only after all checks succeed.
The initial metadata attempt correctly rejected an author Help field longer
than its 4,000-character limit, after its native cases passed. The Help was
shortened, the unchanged check rerun and the final bundle produced. Initial
and corrected evidence are both preserved; no guard was weakened.

Memory cases use GCC AddressSanitizer/UBSan and PCFC11 O0/O1. The host cannot
provide LeakSanitizer's required /proc task inspection, so detect_leaks=0 is
used. Allocation/cleanup counters and active-root checks remain enabled.
Do not describe this as successful LeakSanitizer qualification.

Unchanged installed SDK
The separate SDK 0.6.3 environment contains no PyCForge, NumPy or PySide6.
It connects to the separately installed 1.3.79 review wheel, discovers the
new profile/forms/Help and authors NumPy and independent array namespaces.
Both starters pass 62 cases and the verify/package/register workflow.
Deterministic packaging, stale-evidence refusal with output preservation,
offline reference behavior and engine-upgrade evidence invalidation pass.
All 553 SDK product files remain byte-identical.

The actual installed 1.3.78 engine refuses the new 0.5 profile and preserves
its working 0.4 registration. The current engine admits the actual historical
0.1, 0.2, 0.3 and 0.4 array packages. The SDK version is deliberately unchanged;
engine-owned catalogues and existing forms carry this additive capability.
See sdk-results.json and tools/qualify_sum_mean_sdk.py for exact evidence.

Documentation and final handoff
The standard documentation verifier passes 32 pages, 82 Python examples
(50 Runtime and 32 Typed), five API examples, six C callers, three supplied
modules and three README examples. The new sum/mean manual example also has
its own passing native test. SDK landscape regeneration matches the source.
The README diff is restricted to the release heading and current NumPy
capabilities/limits. Its exact five links, banner widths and renderer pass.

Final wheel/sdist validation, strict twine checks, exact-wheel installed
navigation and production BuildService checks are recorded in the external
PyCForge-1.3.79-Release-Report.txt after the build is frozen. The complete source
archive additionally carries the closed final receipts. This in-distribution
record intentionally does not claim tests on artifacts not yet built.

Preservation and remaining work
All 4,746 paths from the preceding complete source archive are retained;
packaging metadata is refreshed for 1.3.79. Old profiles, numerical packages,
roadmaps, evidence and specifications retain their historical context.
The bundled compiler, math and JSON archives remain byte-identical.

CC reports that 1.3.78's quick function navigation is much better. That is
user-reported Windows feedback. No native Windows desktop/compiler/GC run was
available here; Linux/offscreen Qt results do not close those platform gates
or the prior Preparing-dialog display follow-up.

The next queue is bounded lowering for large literal-generated functions,
then the owned two-dimensional representation and further qualified reductions.
Preserve evaluation order, failure timing, ownership and source diagnostics.
There is no new benchmark, speedup claim or broader C11/NumPy compatibility
claim in this increment. See ROADMAP.md and SDK Framework 0.10 for the plan
and full numerical contract.
