On this page

1.5.0 release qualification

This report records evidence for the typed contiguous foundation. It does not expand the release into any 1.6.0 decomposition or sparse work.

Supported target and configuration results

Target/configuration	Result	Evidence
Windows x86-64, FPC 3.2.2, normal	841/841 tests passed	Full `TestRunner`
Windows x86-64, FPC 3.2.2, `-O3`	841/841 tests passed	Optimised full `TestRunner`
Windows x86-64, FPC 3.2.2, `-O2 -Criot -gh -gl`	841/841 tests; 0 unfreed blocks	Runtime and heap-checked full `TestRunner`
Windows i386, FPC 3.2.2, `-O2`	841/841 tests passed	Optimised full `TestRunner`
Lazarus package 1.5	Passed	Clean-config x86-64 and i386 builds
Published examples	Passed	All 16 compiled and ran
Searchable documentation	Passed	31 pages, 16 indexed examples, 54 new-symbol checks
Clean extracted archive	Passed	Checksum generated; typed solve compiled and ran offline

The repository CI repeats normal Linux x86-64 and Windows x86-64 suites, examples, documentation checks, benchmark compilation, clean-archive quick starts, both Lazarus package builds, and the optimised i386 suite. The publication run completed successfully on PR #9: CI run #92 for commit ddf7af23f15ef53940e9a0a0be3ae39fc861ae7e.

Numerical evidence

Reference matrix products cover real and complex values, single and double precision, 3 x 5 by 5 x 1 odd shapes, (0 x 3)(3 x 2) empty shapes, overlapping destination aliases, and mixed 1e200/1e-200 factors with a representable result.
LU solve tests cover a published 3-by-3 reference solution, multiple right-hand sides, factor reuse, pivot diagnostics, singularity, finite-input validation, and a normalized residual below 1e-14 for the double fixture.
Cholesky tests cover real SPD, tiny-scale SPD, and complex Hermitian positive-definite systems.
Single tests use tolerances on the order of 2e-6; double tests use 1e-12 to 1e-14 according to operation and residual scale.
The scalar special-function budgets and fixtures remain published in MathBase.md and TestMathBase.pas; unsupported families remain visible in the capability inventory.

Performance evidence

On the local Windows x86-64 qualification host (Intel64 family 6 model 141, Windows NT 10.0.26200), an FPC 3.2.2 -O3 benchmark run recorded:

Workload	Elapsed
Compatibility `192 x 192` dense product	15 ms
Typed `(127 x 129)(129 x 65)` product	32 ms

These are reproducibility data, not a cross-library speed claim. The typed benchmark uses deterministic trigonometric inputs and reports a checksum.

Dependency and licence audit

The new stable units import only SysUtils, Math, existing mathlib-fp units, and the Free Pascal RTL/FCL. A source scan found no external declarations, dynamic-library loading, process execution, sockets, or HTTP access in the new matrix, kernel, solver, or single-complex implementation. All new source, tests, examples, scripts, and documentation are covered by the repository MIT licence.

Maturity changes and known gaps

Typed dense storage, arithmetic, LU solve, and Cholesky solve graduate as stable for their documented finite-input contracts. No existing API changes maturity and no API is deprecated.

Known gaps are explicit: single-complex elementary functions beyond the kernel arithmetic subset; Bessel, elliptic, and exponential-integral scalar families; least squares and broader decompositions; typed sparse storage/solvers; parallel/SIMD typed kernels. These do not become 1.5 features implicitly.