mathlib-fp1.9.6
On this page

mathlib-fp 1.9.1 qualification

Completion-gate evidence

Gate Evidence
Defect triage Read-only issue triage on 2026-08-02 found no open repository reports against 1.9.0. Output-aware guide execution exposed one confirmed seeded-bootstrap defect; the correction and permanent regression are described below.
Five-minute first use The README links the exact release page and direct .tar.gz/.zip downloads. Its one-screen program is extracted, compiled with only -Fusrc, executed, and required to print P(Z <= 1.96) = 0.975002 in clean-archive CI.
Versioned web/offline/repository agreement docs/versions.json identifies 1.9.1 as current and 1.9.0 as preserved. documentation.yml rebuilds 1.9.0 from tag v1.9.0, builds 1.9.1 from the release tag, deploys both, and produces the deterministic offline ZIP/checksum from that same site tree. Every HTML page and release.json carries its release identity.
Published qualification release-qualification.yml creates a checksummed archive with git archive, extracts it into an isolated directory on Linux and Windows, runs the configurations below without network access, and publishes JSON plus per-gate logs. The Windows archive job also builds the Lazarus package.
API freeze test_api_snapshot.py and check_docs.py compare all source interface sections with the unchanged schema-2 1.9.0 baseline: 2,880 exact owner/signature-aware declarations and 281 required public names.

Confirmed correction and regression

The seeded bootstrap overload used the low bits of a 32-bit LCG followed by mod Length(Data). For a length-eight input, those low bits cycle through all eight indices, so each nominal resample was a permutation and every mean was identical. The 95% percentile interval in the Stats guide therefore printed [3.2375, 3.2375] instead of expressing sampling uncertainty.

The private implementation now uses TLocalRandom.NextInteger, the shared xoshiro256** explicit-state generator and rejection-sampled bounded-integer kernel. TestSeededBootstrapPowerOfTwoSampleVaries verifies that 64 resample means vary and that the seeded 2,000-iteration interval brackets the sample mean. Existing tests retain exact same-seed replay and verify that RandSeed is unchanged. The checked guide result is [2.7747, 3.7375] for seed 2026. No public declaration, seed parameter, ownership rule, or global-state contract changed.

Output-aware documentation evidence

check_doc_examples.py inventories all 225 Pascal fences, compiles and runs the 15 self-contained programs, and verifies 13 adjacent output contracts. Every runnable program that calls Write/WriteLn must declare either exact output or ordered required fragments; a zero exit code alone does not pass.

examples/output-contracts.json additionally checks numerical values and statuses for examples 15, 16, 22, and 23, then requires their exact final success marker. This prevents a guide from passing after printing a limit or failure status on an otherwise successful process exit. check_built_docs.py then verifies every generated page's release metadata, search-index target, local file link, and HTML anchor. The publishing rehearsal checked 123 landing/release pages across the two version paths.

Reproducible release commands

The cross-platform driver is intended to run inside an extracted source archive:


python tools/qualify_release.py --release 1.9.1 --compiler fpc \

  --lazbuild lazbuild

It records one JSON entry and a complete log for each gate:

  • normal -B tests;
  • optimised -B -O3 tests;
  • checked/heap-traced -Ci -Cr -Co -Ct -Sa -gl -gh tests, including the 0 unfreed memory blocks marker;
  • all 24 examples plus the four output/status/final-marker contracts;
  • API snapshot, documentation unit/static/execution checks, searchable HTML, and offline ZIP/checksum generation;
  • the Lazarus package where lazbuild is available;
  • the representative -O3 benchmark and its bounded-allocation failures.

Observed release matrix

Path Result
Win64 normal 931 tests, 0 errors, 0 failures
Win64 -O3 931 tests, 0 errors, 0 failures
Win64 checked/heap-traced 931 tests, 0 errors, 0 failures; 295,621 blocks allocated/freed and 0 unfreed
Examples and checked output contracts All 24 compile and run; all four release-facing output/status/final-marker contracts pass
Documentation/API snapshot 63 pages; 15 runnable programs and 13 output contracts pass; all 2,880 declarations remain unchanged
Offline documentation 63 searchable pages; deterministic ZIP and SHA-256 generated locally
Lazarus package Version 1.9.1 package builds with Lazarus 4.8/FPC 3.2.2 on Win64
Representative benchmark Compiles/runs at -O3; 100,000-entry sparse and 200,000-dimensional matrix-free cases retain linear storage, allocate no dense-shape elements, and keep warmed heap growth at zero below the 65,536-byte failure ceiling
Clean source archive A disposable prospective-tree commit was archived and SHA-256 verified, extracted in isolation, and passed all 68 Win64 gates; release-qualification.yml repeats this on the exact release tag for Linux and Windows and publishes its authoritative checksum, JSON, and logs

The local Win64 driver completed all 68 recorded gates on 2026-08-02 using FPC 3.2.2, Lazarus 4.8, and Python 3.13.5 on Windows 11 x86-64. Local timing is not a performance threshold. The exact tagged Linux/Windows archive jobs, not this configured checkout, are the final portability and packaging evidence. An independent publishing rehearsal extracted docs/ from tag v1.9.0, built its 59 historical pages beside the 63 current pages, verified both release.json identities and the root's current/older links, and generated the combined offline archive plus checksum. The built-site checker verified all 123 HTML pages and their local links/anchors. The disposable rehearsal tree was then removed.

The clean candidate archive included only Git-tracked and prospective release files, not ignored compiler output or maintainer configuration. Its temporary repository, archive, extraction, and generated results were removed after the 68-gate run. The tag-generated checksums published by CI, rather than a mutable pre-commit candidate digest, identify the release artifacts.

Dependencies, compatibility, and known limitations

The normal library, tests, examples, generated documentation, and offline archive need only FPC 3.2.2 and standard RTL/FCL units. Documentation generation uses Python's standard library. The release paths do not download numerical packages, load a foreign numerical DLL, contact a service, or require a licence key.

The stable capability boundary and limitations remain those published for 1.9.0 in CAPABILITIES.md: no new algorithm family or public convenience surface is included. In particular, none of the planned 1.9.2 beginner-learning-path or later 1.9.x/2.0 convergence work is part of this release.