Release preparation and validation record, not a claim that all platform gates are complete.

Release Preparation¶

Current release candidate: Brecon 1.0.0rc4, tag v1.0.0rc4. The owner authorized GitHub publication and approved the PyPI environment on 2026-10-28. The previous-name v1.0.0rc3 remains immutable. Check GitHub Releases for published assets; source tags and unpublished drafts do not imply a public binary release. Stable 1.0 promotion follows a feature freeze and verification of the actual release artifacts. Native macOS/Windows/Linux development packaging is available in source; public signed installers and broader platform coverage remain later steps.

Known RC4 plate limitation: independent post-release checks found raw Rect moment/shear recovery inconsistent with verified plate laws in PyNite 3.0/3.2. Quad moments are preferable for RC4; Quad transverse shear is also unverified and must not be relied on. Development disables Quad shear plots/data. The Rect correction and independent opening/Rect benchmarks are in 1.0.0rc5.dev0 source development; RC4 assets are unchanged. Non-unit Rect bending modifiers are rejected in development because the supported engine ignores them in bending stiffness. See the plate guide for scope and benchmarks.

Publication status (2026-10-28): Brecon RC4 is published on GitHub and PyPI. Workflow 40713196939 verified the tagged assets and published the same files through Trusted Publishing. The PyPI hashes match SHA256SUMS, and a fresh public-registry installation reports Brecon 1.0.0rc4 from both launch commands.

PyniteGUI RC3 remains published as a GitHub prerelease and on PyPI; its hashes match the sealed checkpoint. Previous tags and assets are unchanged. Stable 1.0 platform and prerelease-feedback gates below remain open.

Build and Verify¶

From the repository root, with uv installed:

uv sync --locked --extra mcp
QT_QPA_PLATFORM=offscreen PYNITEGUI_NO_WEBENGINE=1 uv run --locked --extra mcp python -B -m unittest discover -s tests -q
uv build --no-sources
uv run --locked python -B scripts/check_dist.py dist/brecon-1.0.0rc4-py3-none-any.whl dist/brecon-1.0.0rc4.tar.gz

The distribution checker audits versions, metadata, bundled runtime assets, licenses and source contents and writes dist/SHA256SUMS for these two artifacts. The wheel must contain the offline viewport/vendor assets and both libraries. The README is also the PyPI long description: use absolute HTTPS image and document URLs and check them in the built metadata, not just GitHub's renderer. Updating the repository README does not update an already uploaded PyPI description; publish a new version for such changes rather than replacing files. The source archive also includes the lockfile, guide, release documents, tests, scripts and README screenshots. Do not attach old-version artifacts.

Install the wheel into a fresh environment outside the checkout. Run the smoke script with that environment's Python from a temporary working directory, not with uv run in the repository (which would test the editable project instead). Example POSIX commands; use the corresponding Scripts/python.exe on Windows:

uv venv --python 3.12 /tmp/pynitegui-rc-base
uv pip install --python /tmp/pynitegui-rc-base/bin/python ./dist/brecon-1.0.0rc4-py3-none-any.whl
cd /tmp
/tmp/pynitegui-rc-base/bin/python /absolute/path/to/PyniteGUI/scripts/smoke_installed.py --version 1.0.0rc4

Repeat in another fresh environment, installing the same wheel with [mcp] and adding --mcp to the smoke command. The script checks that imports are not editable, bundled assets are present, all examples save/open/analyze, exports work and native windows initialize; the MCP variant also exercises a real local SDK client and server lifecycle. It is not a visible desktop/WebGL smoke test.

For repeatable visible checks, use that MCP environment's Python in a desktop session with WebEngine enabled:

cd /tmp
/tmp/pynitegui-rc-mcp/bin/python /absolute/path/to/PyniteGUI/scripts/smoke_desktop.py --output /tmp/pynitegui-desktop-qa --hold

Use a new output directory for each run. This script uses disposable settings, real native widgets and the actual renderer, with QtTest/application commands and supplied file paths. Its screenshots and JSON diagnostics are evidence of those automated workflows, not human mouse/keyboard or OS file-picker checks. With --hold, complete the remaining manual checks in the open test window: Save As and reopen through the real pickers, F5 analysis, and File > Export Results > Print Results. Record the outcome before ticking the macOS manual gate.

Run tests/viewport3d.spec.cjs with a separately installed Playwright/Chromium as described in GUIDE.md. These browser tests supplement, not replace, native desktop checks. Check the built wheel's 3D viewport on each target platform. Set VIEWPORT_ASSETS to the installed wheel's pynitegui/qt/viewport3d directory to serve its assets instead of the source checkout during browser testing.

Stable 1.0 Scope¶

Elastic 2D/3D frames and bilateral axial-only trusses, plus frame-only P-Delta within GUIDE.md restrictions. Surface meshing, plate analysis and coupled Quad slabs remain explicitly experimental. Nonlinear/plastic, modal, unilateral and reinforced-concrete workflows are deferred. Remaining UI roadmap items do not block this scope.

Historical RC3 Checklist and Future Publication Gates¶

• Project citation guidance and GitHub citation metadata added without a DOI.
• For each release, review CITATION.md/CITATION.cff and preserve exact version and commit links. Consider Zenodo concept/version DOI archiving for stable 1.0; archive later releases as new versions of the same record.
• RC3 full suite passes with MCP wire tests enabled, not skipped.
• RC3 wheel/source archive audits and base/MCP installed-wheel smoke checks pass.
• Native Linux visible startup, 2D/3D editing, save/open, analysis, reports and MCP checked.
• Repeat installation and automated visible workflows on macOS; record OS/Python/Qt/renderer versions.
• Human macOS file-picker, F5 shortcut and interactive print-dialog checks confirmed.
• Windows installation and visible workflows checked, or explicitly labelled unverified.
• RC3 release notes and limitations reviewed; artifact hashes generated in SHA256SUMS.
• User approves the consolidation checkpoint: push changes and prepare an unpublished RC3 draft with assets.
• RC3 GitHub/PyPI publication confirmed on 2026-10-28; its tag and assets remain unchanged.
• Historical RC2 GitHub/PyPI publication completes; its artifacts and tag remain unchanged.

Upload the wheel, source distribution and SHA256SUMS to the approved GitHub prerelease, using CHANGELOG.md for notes. Tag the verified commit, not a moving branch. GitHub also provides source archives for the tag. Preparation scripts do not upload to PyPI; use the manual publishing workflow below. For stable 1.0.0, update version/lockfile/notes, rebuild and repeat validation rather than renaming RC artifacts.

uv Tools and PyPI¶

uv installs tools from Python package indexes; PyPI is the intended public distribution channel. The brecon entry point is canonical; pynitegui is a compatibility alias. Use the publication status above to confirm availability. A separate uv-specific package or registry submission is not needed.

After stable 1.0 is published to PyPI, users can run:

uv tool install --python 3.12 brecon
brecon
uv tool upgrade brecon

For optional MCP support, install "brecon[mcp]" instead. An RC published to PyPI would require an explicit version, for example "brecon==1.0.0rc4". These registry commands describe the publication plan; no PyPI availability is claimed by this checklist. Local-wheel installation above works before upload.

Before upload, confirm ownership/availability of the PyPI project name and set up a PyPI account and publishing credentials, preferably GitHub Actions Trusted Publishing. Upload only the reviewed wheel and source archive, with explicit paths, after final approval. TestPyPI can be used as a separate dry run; it is not a substitute for testing the artifacts locally. Never commit API tokens.

One-time Trusted Publishing setup: sign into PyPI, enable account 2FA if needed, and open Publishing. Add a pending GitHub publisher with these exact values:

Brecon needs a new pending publisher, separate from the existing pynitegui project. The owner confirmed configuring it on 2026-10-28; the first RC4 upload succeeded on that date. Keep the old project/releases for existing users. Its package name is not reserved until the first successful publish. See the PyPI pending-publisher guide.

A pending publisher creates the project on its first successful upload; it does not reserve the name. If you already own the project, add this publisher under that project's Publishing settings instead.

In GitHub repository Settings > Environments, create pypi, add yourself as a required reviewer, allow your own review if you will dispatch the workflow, and restrict deployment branches to master. No PyPI API-token secret is needed. These account/environment changes are owner setup, not performed by adding the workflow.

Publish an approved release: first publish the GitHub draft with its wheel, source archive and SHA256SUMS attached. Then open Actions > Publish Brecon to PyPI > Run workflow, choose master, and enter its tag (for example v1.0.0rc4, once prepared). Review and approve the pypi environment job. This uploads to real PyPI; it is not a dry run. The workflow derives the distribution name from the tag metadata and accepts Brecon only. Old PyniteGUI RC2/RC3 tags are rejected before assets are downloaded; deleting old releases is unnecessary. The rename itself did not publish or replace RC3 assets; RC4 was released separately after validation.

The workflow is manual only. Its read-only verification job checks out the requested tag, rejects a version mismatch or draft release, downloads the exact GitHub assets, checks both hashes, and audits them against the tagged source. The separate environment-gated publication job has OIDC permission and uploads those verified files without rebuilding. It does not run on pushes or tags. Do not move existing release tags or replace reviewed release assets to retry.

References: uv tools and building and publishing.

Post-RC4 P-Delta Support-Movement Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. Public RC4 assets and file formats (planar v29 / spatial v36) are unchanged. This extends an existing analysis method; the hold on new iterative/unilateral, modal and plastic solvers remains in place.

• Remove the blanket nonzero-motion P-Delta guard after independently validating the existing sparse affine solver. Every geometric update includes the full prescribed displacement in its right-hand side. Retain geometric-stiffness convergence, positive effective free stiffness, physical equation checks and independent global force balance before publication. Update progress/error context to cover prescribed as well as inclined support constraints.
• Prescribed global translations/rotations and inclined-normal motions retain their model-level values in every combination, including zero applied loads. Support shortening can itself produce axial preload and second-order response. Ground springs keep zero rest positions. Retain exclusions for trusses, surfaces, released translations and spatial rotational releases; supported planar moment releases still transfer zero released-end moment.
• Independent references import neither application nor PyNite: solve the exact elastic beam-column state-transfer equation for compression/tension, base translation/rotation, differential support movement, imposed axial shortening, biaxial bending, uniform loading and a ground spring. Check mesh convergence at 1/4/24 segments, nodal/reaction fields and approximate interior deformation. A separate six-DOF energy reference integrates cubic Hermite curvature/slope products and uses KKT constraints plus a scalar axial root. Check arbitrary spatial normals, additional rigid rotation, six base motions, springs, planar conversion, signed/zero combinations, released chord response, rigid translation, shortening-induced buckling and forced nonconvergence rejection.
• Distinguish the strain-derived geometric preload from external axial force: the engine's axial geometric term gives N + N^2/(EA) = applied axial force for a uniform force-controlled axial state. This is the existing small-strain geometric-stiffness convention, not a new finite-strain constitutive law. P-Delta interior member recovery remains approximate; mesh refinement and the existing force-only global equilibrium convention remain explicit in GUIDE.
• Add P-Delta > Column - Imposed Base Motion, including Browse Examples: 480 in column in eight segments, base DX +0.03 in / RZ +0.0007 rad, gravity -60 kip and lateral +1 kip. Combined nodal tip DX is 3.5186619 in versus continuum 3.5186638 in; Gravity gives -0.47371345 in. Base motion only gives -0.306 in and zero member forces. Add the actual native screenshot, staged edit/method undo, spawned worker, unit conversion, snapshot invalidation and authenticated MCP edit/solve/zero-state result coverage. MCP discovery states the supported method scope and unchanged-per-combination motion rule.
• Validation: 1,040 tests, no skips, 295.656 s, including the optional MCP wire suite. The final installed wheel passes 63 focused tests on PyNite 3.0.0 / Qt 6.11.2 (4.774 s) and PyNite 3.2.1 / Qt 6.12.0 (4.823 s). All 32 examples save/reopen/analyze/export and native windows initialize; authenticated installed MCP discovery/edit/undo also passes. The installed Python files match the wheel bytes. Strict docs and local links cover 55 HTML pages; browser QA verifies the new native screenshot, existing imagery, search, subpath, dark mode and layout. Distribution metadata/assets/licenses/source audit passes. Test ownership covers 158 classes / nine groups. Manual platform/signing gates remain unchanged.

Post-RC4 Quad Shear Feasibility Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. Test/documentation changes only; production solver, snapshots, UI, file formats and published RC4 assets are unchanged. Quad transverse shear remains unavailable in results, CSV and MCP.

• Add a reproducible comparison of raw Quad shear, the divergence of each element's bilinear moment field, and a quadratic fit to neighbouring element- centre moments. The fit uses only complete 3 × 3 rectangular patches and reports no value at incomplete patches. Manufactured quadratic tensors check derivatives, signed/zero scaling and coordinate conditioning. Missing-centre tests prevent crossing a hole or extrapolating to an edge; nonuniform and invalid grids are rejected.
• Solve 120 × 180 × 0.6 in isotropic rectangles under -1e-7 kip/in², E=3600 ksi: simply supported Navier plate and coupled slab with negligible-stiffness edge beam (E=1e-6 ksi), plus nu=0 cantilevers clamped in either local direction. Three meshes (24/96/384 Quads), XY/XZ/YZ and Service/-2 Reverse/Zero produce 36 model solves / 108 reported combination records per engine version. Every combination passes the existing six-component global equilibrium gate; no application guard or tolerance is relaxed.
• Check the independent Navier shear series at 160 versus 320 terms over the fine-mesh comparison region; component changes stay below 0.02%. Compare methods on identical complete interior patches, excluding centres within 20 in of the outer boundary. At the fine mesh, raw shear has 58.41% vector L2 error and own-element moment differentiation 22.73%. The neighbouring fit converges from 2.829% → 0.667% → 0.148%; its largest fine component error is 0.170%. All three methods pass the simple cantilever cuts, demonstrating why those cuts alone cannot validate general plate shear.
• Locked PyNite 3.0.0 and installed 3.2.1 reproduce the same bounded comparison and pass all 108 reported records. The fit remains a test-only candidate: boundary/opening recovery, irregular meshes, material/stiffness jumps, thick plates, curved shells, peaks and rotational interfaces remain unvalidated. It smooths across neighbours and does not replace unsmoothed production probes. Preserve the original unchecked exposure task and add explicit integration prerequisites to TODO.md.
• Reproduce JSON and the comparison figure with scripts/benchmark_quad_shear.py --output /tmp/quad-shear.json --figure /tmp/quad-shear.png. The report's checks_passed refers only to the bounded research checks; individual raw/candidate error metrics remain explicit.
• Validation: 143 surface/docs/tooling tests, no skips, 115.815 s. An installed locked-engine wheel passes six new tests, 56.120 s; installed PyNite 3.2.1 / Qt 6.12 passes 26 recovery/surface/probe tests, 63.090 s. Ownership covers 157 classes / nine groups. Strict docs and local links cover 55 HTML pages; browser QA checks the new figure, existing workflows, search, subpath, dark mode and layout. Distribution audit passes; the full frame/MCP suite is not repeated for this test/documentation- only change. The previous production checkpoint remains recorded below.

Post-RC4 Imposed Support-Movement Checkpoint¶

Development 1.0.0rc5.dev0, 2026-10-29; public RC4 assets remain unchanged.

• Add signed global support translations/rotations and inclined-normal displacement to the existing linear solve. Use canonical inches/radians and display-unit-aware staged Properties, numerical tables and explicit bulk opt-ins. Prescriptions are model-level and unchanged in every combination, including zero applied loads.
• Solve with an affine physical lift and free-coordinate elimination. Preserve global spring rest positions at zero; independently recover physical reactions, released-end motions and all-combination free/global equilibrium. Support exact redundant constraints only when their prescribed motions agree. Shared shell edge constraints remain active; incompatible prescriptions reject analysis.
• Cover linear 2D/3D frames/trusses and supported releases, explicit shared surface joints, support copying/collision identity, conversion, snapshot invalidation, process-worker transport, definition reports and paginated unit-labelled MCP movement metadata alongside normal/unused-rotation metadata.
• Add Beam - Support Settlement and Rotation, with right DY -0.1 in and RZ +0.001 rad. Its zero-load Settlement only case has N1 FY +16.433333 kip / MZ +862.75 kip-in and N2 FY -16.433333 kip / MZ +1109.25 kip-in. Record a reproducible native offscreen illustration and explicit load-factor semantics.
• Advance planar saves to v29 and spatial saves to v36. Current movement fields are required; older files migrate zero motion without changing supports. Reject malformed values, nonzero old-schema motion and free-direction prescriptions.
• Nonzero support movements reject P-Delta pending independent second-order recovery checks. No case-specific prescriptions, moving spring rest positions, staged-construction or plastic/nonlinear connection method is introduced.

Validation of the support-movement extension:

• Full locked source suite with optional MCP: 1,022 tests, no skips, 254.851 s; 156 classes / nine groups. Focused model/editor/MCP/table follow-up: 61 tests, no skips, 3.345 s.
• The installed wheel, outside the checkout, passes 100 focused tests on each of PyNite 3.0.0 / Qt 6.11.2 (7.302 s) and PyNite 3.2.1 / Qt 6.12.0 (6.681 s), without skips. Installed smoke passes all 31 examples through save/open/analysis/export and native initialization, plus real authenticated MCP discovery/edit/undo.
• The full suite includes an authenticated HTTP client that edits a settlement, runs the real process worker, reads unchanged prescribed motion in the zero-load combination, and verifies invalid-edit rollback without changing the revision. Independent references cover fixed-ended beam stiffness/Hermite fields, 2D/3D normal compliance, ground springs, rigid-body motions, trusses, detached member-side translations and fully released joint rotations without false torques.
• Strict documentation/local links pass across 55 HTML pages; documentation browser checks include the inspected native settlement illustration, search, subpath, dark mode and layout. Offline viewport interaction/support/hover, graphics-resource and context-loss checks pass. Wheel/source metadata, runtime assets, notices and source-content audit pass. No public release or additional human OS/hardware/signing validation is implied.

The following sections record earlier development checkpoints and their historical scope/validation; the imposed-motion exclusions there are superseded by this linear-only extension.

Post-RC4 General Inclined-Support Checkpoint¶

Development 1.0.0rc5.dev0, 2026-10-29; public RC4 assets remain unchanged. This extends the planar checkpoint below across existing model and analysis types.

• Spatial v34 stores required normal azimuth/elevation. Two tangent translations remain free; optional global restraint flags add exact homogeneous constraints. Exact redundancy is deduplicated; nearly dependent constraints and springs on fully constrained directions reject the model. Planar v27 keeps tangent angles.
• The sparse constrained solve includes trusses, linear end releases and shared coupled-surface nodes. Physical free equations use the unaugmented matrix; unused-rotation gauges never become ground torques. Existing shell drilling, release/surface and independent equilibrium gates remain active.
• Existing frame P-Delta starts each combination elastically and updates geometric stiffness until physical equilibrium closes, with at most 30 updates and positive constrained effective stiffness required. This closes axial-strain coupling beyond the engine's ordinary two-step approximation. Global P-Delta checks remain force- only; no truss/surface/translation-release/spatial-rotation-release P-Delta scope is added. Settlements and plastic/nonlinear connection methods remain separate.
• Staged native Properties, tables, bulk edits, global support copying, 2D-to-3D conversion, normal-facing offline viewport symbols/picking, orthographic reports and six-component paginated MCP reaction metadata use the same definitions. Released-rotation metadata is retained alongside inclined reaction metadata.
• Add 3D Cantilever - Oblique Roller Normal and a native Properties capture. Independent cantilever compliance and dense P-Delta references cover all six motions/reactions, poles, oblique normals, signed/zero combinations, spring axes, additional constraints/redundancy, releases/trusses, conversion, coupled nodes, buckling/nonconvergence and process-worker snapshot transport.

Validation of this extension:

• Full locked source suite with MCP: 1,009 tests, no skips, 252.652 s; 154 classes / nine groups. Final schema/editor/mechanics follow-up: 40 tests, no skips, 4.047 s.
• Installed-wheel compatibility outside the checkout: 76 focused tests each on PyNite 3.0.0 / Qt 6.11.2 (5.459 s) and PyNite 3.2.1 / Qt 6.12.0 (5.378 s), with no skips. The installed locked runtime also passes all 30 examples through save/open/analysis/export/native initialization and a real authenticated MCP discovery/edit/undo session.
• Wheel/source metadata, bundled assets, notices and source-content audit pass. Offline viewport checks cover normal-facing geometry, roller/pick identity and actual example hover alongside existing interaction/result/context-loss checks. Focus-resource tests explicitly render pending texture uploads before retaining their exact leak assertions. The native offscreen Properties capture was inspected.
• Strict documentation/local links cover 55 HTML pages. Browser documentation checks pass imagery (including spatial normal controls), search, subpath, dark mode and layout. Public releases and manual OS/hardware/signing gates are unchanged.

The following section records the earlier, narrower checkpoint and its historical validation results; its exclusions are superseded by this extension.

Post-RC4 Planar Inclined-Roller Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. This extends the existing linear frame adapter; new solver methods remain on hold and public RC4 assets are unchanged.

• Add an inclined node support with a signed global tangent angle in degrees. Its normal translation is fixed; tangent translation and RZ stay free. Assemble sparse node-local free coordinates without penalty stiffness or artificial grounding. Verify constrained rigid-body/stiffness stability (including the sparse spectral branch and pre-cancellation roundoff bound), free physical equations and all-combination independent global force/moment equilibrium.
• Store physical global displacements for ordinary engine member recovery. Independently recover rigid normal force and global bilateral spring reactions; ordinary node tables retain their global totals. Native Properties shows the signed rigid normal force; paginated MCP node reads provide separate rigid/spring vectors, normal force, explicit units and chosen normal direction.
• Add staged Properties, Model Tables and bulk edits, rotated native/report symbols, tree/tooltips, definition reports and an editable Beam - Inclined Roller example. Its 120 in pin/30-degree roller beam under -0.01 kip/in gives Service N2 FX -0.346410 / FY +0.6 kip, rigid normal force +0.692820 kip; N1 balances FX and carries FY +0.6 kip. Reversed combinations flip signs. The guide includes an inspected native offscreen capture; it is not a human OS-dialog/hardware validation claim.
• Planar saves advance to v27, with required node support_angle. Older versions migrate to zero angle and retain their original supports. Reject malformed angles, old-version inclined inputs and missing current fields. Spatial stays v32 and omits the inactive inherited angle field; nonzero spatial angles are rejected. Plate files remain v3.
• Require unreleased planar linear frame members. Reject P-Delta, trusses, released/spatial frames and spatial conversion before solving or changing the source. Copied support orientations remain global, with explicit collision handling; changing inclination invalidates snapshots while unit-only changes retain them. Real process-worker, metadata pagination/unit conversion, undo, input precision and invalid-batch rollback regressions are included.
• Independent axial/Euler–Bernoulli cantilever and pin/roller references cover oblique/reversed beams, member fields, signed/zero combinations, global springs, nonuniform subdivisions and deliberately corrupted recovery. Sparse failure and near-mechanism cases reject results. No nonlinear connection, imposed movement or additional solver method is enabled.
• Full locked source suite with MCP: 994 tests, no skips, 247.393 s; 152 classes / nine groups. Installed locked PyNite 3.0.0 / Qt 6.11.2: 52 focused tests, no skips, 4.095 s. A second installed wheel runtime using PyNite 3.2.1 / Qt 6.12.0 passes the same 52 tests, no skips, 5.160 s. The final smoke/UI/inclined-support follow-up passes 192 tests, no skips, 72.793 s, including reaction-label updates after solving/combination changes without losing unapplied edits. Installed smoke verifies 29 examples through save/open/analysis/ export and native initialization, plus real authenticated MCP discovery/edit/undo. Strict documentation/local links cover 55 HTML pages; browser QA verifies support-page imagery, search, subpath, dark mode and layout. Public releases and the manual platform/signing gates remain unchanged.

Post-RC4 Inclined-Support Feasibility Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. This is a test-only feasibility study. No production adapter, UI, file format, MCP schema, public tag or RC4 asset changes.

• Demonstrate exact homogeneous planar support constraints through a free coordinate basis, without penalties or artificial grounding. Scale translational and rotational coordinates, reject dependent/nearly dependent constraints, require positive stiffness on every physical free mode, and verify constraint and free-equation residuals before recovering global rigid reactions. Carry a pre-cancellation matrix-product roundoff bound through diagonal scaling; reduced normalization alone initially missed a near-mechanism. Reject that unresolved mode rather than accepting its small positive numerical stiffness.
• Prepare actual engine member stiffness and equivalent nodal loads (including fixed-end vectors) on a disposable all-clamped copy. Discard its displacements; solve the test-only constrained system without creating AnalysisResults. Compare with independent axial/Euler–Bernoulli cantilever flexibility and statically determined pin/roller beam fields. No application guard is bypassed.
• The reproducible report covers 153 cases: four beam orientations, nine roller angles, three subdivisions, three global-spring setups, signed/zero combinations. Maximum reference errors: 1.01e-15 in translation, 3.94e-17 rad rotation, 3.29e-13 kip force and 1.80e-12 kip-in moment. Independent global balance residuals stay below 4.45e-14 kip / 2.66e-12 kip-in. Six unsupported/dependent configurations are rejected, including zero-load mechanisms; on a horizontal beam, a pin plus a horizontal-normal roller does not prevent rigid rotation.
• Tests also compare a separate Lagrange-multiplier system, individual nodal and uniform member load patterns, coordinate/normal-sign/scaling invariance and existing global-axis production supports. Keep global bilateral spring axes unchanged and separate their external -k*d reactions from rigid reactions. Quadratic-energy identities refer only to the equivalent nodal system, not total distributed-load strain-energy recovery.
• Production integration remains a separate unchecked TODO: sparse node-local transformations, constraint-aware stability/rigid-body checks, member recovery, reaction metadata, persistence/migration, staged UI and MCP, drawing/reporting, conversion/transform semantics and worker/snapshot validation. Proposed angle convention: roller tangent counterclockwise from global +X, in degrees. Inclined supports are still unavailable; P-Delta, spatial/released frames and imposed movements are outside this checkpoint. Solver-method work stays on hold.
• Locked PyNite 3.0.0 / Qt 6.11.2: 256 focused engineering/support tests, no skips, 4.950 s. Installed PyNite 3.2.1 / Qt 6.12.0: 11 new tests, no skips, 0.187 s. Docs/tooling: 16 tests, 0.487 s. Ownership checks cover 150 classes / nine groups. Strict docs/local links cover 55 HTML pages; browser checks pass for images, search, subpath, dark mode and layout. Reproduce the bounded numeric report with scripts/benchmark_inclined_supports.py.

Post-RC4 Nonuniform Membrane Validation Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. This is a test/documentation checkpoint; no production solver, UI, file format, public tag or RC4 artifact changes.

• Add two independent zero-body-force Airy plane-stress references: pure in-plane bending and mixed nonuniform normal/shear stress on a 120 × 80 × 0.8 in rectangle. Integrate edge tractions consistently, retaining the exact force/end-moment resultants. Only three in-plane scalar restraints remove rigid motions; normal-displacement edge supports suppress bending. An off-plane corner frame anchor carries zero load; no solved displacement is prescribed.
• Solve the unchanged application path on 6/24/96/384 Quad meshes in XY/XZ/YZ, with Service, -2 Reverse and Zero combinations: 24 solves / 72 result records. Every combination passes the existing independent six-component equilibrium gate. Displacement and symmetric stress-tensor L2 errors use 3 × 3 Gauss integration; compare external work/2 with independent closed-form strain energy. All three errors decrease monotonically. Fine-mesh Service errors are:

• Independently check constitutive compatibility, stress divergence, energy and rigid-mode gauges. Cover translated origins, alternate Poisson ratio, canonical unit metadata, file round-trips and unchanged input definitions. Confirm that nonzero continuum spin coexists with zero independent nodal drilling rotation; these membrane checks do not validate frame–shell rotational interfaces.
• Locked PyNite 3.0.0 / Qt 6.11.2: 121 surface tests, no skips, 58.506 s. Installed PyNite 3.2.1 / Qt 6.12.0: the five new tests pass without skips in 32.869 s. Docs/tooling: 16 tests, 0.474 s; ownership checks cover 149 classes / nine groups. Strict docs/local links cover 55 HTML pages; browser checks pass for images, search, GitHub Pages subpath, dark mode and desktop layout. Reproduce the report/figure with scripts/benchmark_nonuniform_membrane.py; the guide includes the convergence plot.
• General wall/diaphragm behaviour, corner stress peaks, drilling treatment and rotational interface validation remain pending. The separate lateral drilling case is still rejected under refinement. Quad shear stays withheld; no stiffness or acceptance tolerance is changed. Solver-method implementation stays on hold.

Post-RC4 Spatial Translational Release Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. Implementation commit 5bef8ad. This extends the existing linear adapter; solver-method work stays on hold and RC4 release artifacts are unchanged.

• Add start/end local DX/DY/DZ releases alongside RX/RY/RZ. Validate boolean flags, forbid releasing axial DX or torsional RX at both ends and internally singular bending patterns, and retain global mechanism checks. Translations never enter the unused-rotation coordinate gauge; unsupported disconnected joint translations remain rejected.
• Recover eliminated member-side motions from the uncondensed linear stiffness and fixed-end vector. Check every recovered released-end force/moment equation, every free physical joint equation and all six global balances for each combination. Independent axial/Euler–Bernoulli flexibility checks cover either end, both planes, partial/varying/point loads, oblique/vertical/reversed axes, roll, signed/zero combinations, springs/other incident members, splitting and transport. Deliberately corrupted recovery must fail before its result is used.
• Expose staged Properties, Model Tables, bulk and atomic MCP editing, viewport labels/deformed member paths, local end translations in Properties/reports and paginated released_end_translations metadata in MCP member reads. Preserve unit- only snapshots, undo and invalid-batch rollback. Extend 2D-to-3D conversion to axial/shear releases with independent source windows and field/reaction parity.
• Spatial files advance to v32; older versions retain connected translations. Reject old-version nonfalse flags, missing required fields and malformed flags. Planar files stay v25 and plate files stay v3. Add 3D Fixed-Guided Beam - Translational End Releases, with an independently checked 204 in sliding-end beam. Recovered Service end DX/DY/DZ are +0.000139324 / -0.00487912 / -0.0975823 in, while both joints have zero translation.
• Full source suite with MCP: 966 tests, no skips, 215.315 s; 148 classes belong to nine groups. Independently installed locked wheel: 58 focused tests, no skips, 3.892 s. Strict docs/local links cover 55 HTML pages; documentation browser QA, offline spatial viewport QA and wheel/source audits pass. A second installed runtime using PyNite 3.2.1 / Qt 6.12.0 passes 23 focused numerical/connection/UI checks, no skips, 2.934 s.
• Native standalone run 40909523162 passes on Windows Server 2025 x86_64, Ubuntu 22.04/glibc 2.35 x86_64 and macOS 15.8.2 arm64, with Python 3.12.16 / Qt 6.11.2 / PyNite 3.0.0. Each relocated app passes seven native check stages covering 28 examples, including actual translational-release worker/control/diagram checks, plate worker, PDF and authenticated MCP. Native screenshots were inspected; downloaded archives match their SHA256SUMS. The guide includes the actual Linux release-control screenshot.
• Released P-Delta, frame/surface coupling, nonlinear connections and translational coordinate gauges remain unavailable. Standalone signing/installers and human platform gates remain open; no new public release or PyPI upload was made.

Post-RC4 Native Windows/Linux Packaging Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. Source commit 15bbfa7; native validation run 40861868669 passes on all three operating systems. These temporary Actions artifacts do not publish a release, create a tag or upload to PyPI. RC4 remains unchanged.

• Build on each native host using Python 3.12.16, locked Qt 6.11.2 / PyNite 3.0.0, PyInstaller 6.22.3 and hooks 2026.9. Pin platform-specific build dependencies, render the existing mark as Windows ICO/macOS ICNS/Linux PNG, record host/architecture/libc and keep the reviewed wheel/source identity. macOS/Linux archives preserve symlinks; Linux also preserves executable modes.
• Verify wheel/source metadata, assets and recipe bytes before freezing. Verify archive SHA-256 before extraction. Launch from a new directory with PYTHONHOME, PYTHONPATH and VIRTUAL_ENV removed and PATH limited to system locations; audit the exact reported executable, bundled imports, notices/source hash, WebEngine helper and version metadata. Retain the macOS ad-hoc signature check.
• All three frozen apps pass the six check stages: 27 examples each through save/open/analyze/export, spawned released-member analysis with hidden All tools Results retained, offline WebEngine/PNG, Qt PDF output, spawned plate analysis/ local inspection/Matplotlib PNG and authenticated MCP schema/edit/solve/results. Saved native-window, viewport and plate images were inspected; downloaded archives match their SHA256SUMS. Retain JSON, images, PDF, dependency/build record and corresponding source with the artifacts. Actions retention is 14 days.
• Native packaging/tooling tests: 27 on each host; Windows omits two explicitly POSIX-only symlink/mode tests. Linux and macOS have no skips. The initial native run caught a missing Linux XCB shape library and case-sensitive lookup after copying the Windows environment; both are fixed, with a Windows regression test. Local packaging/tooling/docs tests, strict documentation/link checks across 55 HTML pages, browser documentation QA and wheel/source audits pass.
• CI uses software graphics, with Xvfb on Linux; it does not validate consumer Windows versions, every Linux distribution, Wayland/GPU drivers, native file pickers or printing to physical devices. Windows/Linux previews are unsigned; macOS remains ad-hoc signed. Public installers, signing/notarization, SmartScreen/Finder quarantine and human platform feedback remain release gates.

Post-RC4 Desktop Packaging Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. This work creates local macOS development artifacts and a manually dispatched artifact-only build workflow; it does not create a release tag, publish a release or upload to PyPI.

• Build an architecture-specific Brecon.app from the reviewed installed wheel and corresponding source archive, with PyInstaller 6.22.3 / hooks 2026.9 and a fresh locked runtime. Reject editable/stale installations, source/recipe mismatches and runtime drift. Include the existing vector mark as the app icon, distribution metadata/licenses, AGPL/citation files, source archive and dependency/ wheel/source-hash provenance. Derive the minimum macOS bound from Python and native wheel tags without claiming manual coverage of that lower bound.
• Archive with symlink-preserving ditto; retain SHA256SUMS and build-info.json. Extract the ZIP to a different location and run the embedded desktop check with Python paths removed and PATH limited to system utilities. Audit resources, symlinks, corresponding-source hash and the ad-hoc signature separately.
• The frozen check covers all 27 examples through save/open/analyze/export, spawned frame/released-member and plate workers, actual offline Qt WebEngine and PNG export, Matplotlib/local plate inspection, Qt PDF output and authenticated MCP schema/edit/background-analysis/result reads. It uses disposable settings and writes a record plus native/plot/PDF evidence for that exact executable.
• The packaging exercise exposed a real native WebEngine/MCP transition crash under both Qt 6.11.2 and 6.12, including the regular installed application. Replace the application-wide Python input filter with a bounded set on known editor widgets/actions, refresh after edits and remove on unlock. The original native reproduction now completes with input/shortcut protections retained.
• Verified locally on macOS 27.2 arm64, Python 3.12.13, locked PyNite 3.0.0 / Qt 6.11.2. The relocated frozen app completes every check, including the plate- to-MCP transition, and passes the bundle/source/signature audit. Native screenshots and plate/PDF exports were inspected. The archive is about 228 MiB; its exact identity is recorded in its SHA256SUMS and embedded build-info.json.
• Full source suite with MCP: 948 tests, no skips, 252.752 s. The independently installed wheel using the exact locked bundle dependencies: 948 tests, no skips, 272.837 s. Test ownership: 147 classes in nine groups. Final focused input-filter/packaging/tooling checks pass. The native crash reproduction also passes with the repaired code under Qt 6.12. Strict documentation/link checks cover 55 HTML pages and browser QA passes. Wheel/source audits include all desktop recipes, scripts, tests and workflow.
• At this earlier checkpoint, Developer ID signing, notarization, Finder/quarantine handling, Intel and Windows/Linux artifacts were unverified. The preview has an ad-hoc signature and must not be presented as a notarized public installer.

Post-RC4 Spatial Rotational Release Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. This checkpoint does not publish RC5 or alter the immutable RC4 artifacts.

• Add local RX/RY/RZ start/end frame releases for Linear elastic spatial models without surfaces, with staged inspector/table/bulk/MCP editing, definition reports, endpoint labels and an inclined biaxial beam example. One RX end may release torsion; releasing both is rejected. Translational releases, released P-Delta and released frame/surface interfaces remain unsupported.
• Compute oblique/rolled unused joint-rotation bases, verify K G = 0 and Gᵀ(P − FER) = 0, and use a numerical coordinate gauge equivalent to eliminating those coordinates. Physical supports/springs are unchanged. Real connection mechanisms, nearly dependent axes and unresisted moments are rejected. Verify every free physical equation and whole-model equilibrium for every combination before accepting a snapshot.
• Recover local member-side end motion from PyNite's uncondensed matrices. Properties shows local RY/RZ end slopes separately from undefined global joint components, which remain n/a in tables/CSV/MCP. MCP exposes page-scoped global unused-rotation directions. Free uniform torsional angles are not unique.
• Spatial files advance to v30 with strict release fields and legacy migration. Explicit 2D-to-3D conversion now preserves rotationally released frames, including Gerber/three-hinged arch response; translational releases still block conversion. There are 27 built-in examples; upstream coverage stays nine of 20.
• Independent references cover biaxial beam moments/deflections/end slopes, fixed-pinned internal hinges, free-tip releases, point torque, signed/zero combinations, roll and oblique/reversed/vertical axes, springs, mixed incident connections and actual mechanisms. Fault injection verifies rejection of an entire snapshot when only a later combination has bad physical equations or reactions. Native installed-wheel checks use the actual background worker and preserve hidden Results in All tools. A visually inspected native screenshot documents the new controls and recovered slopes.
• Full locked suite with MCP: 939 tests, no skips, 249.353 s. The independently installed wheel with PyNite 3.2 / Qt 6.12 passes 939 tests, no skips, 277.210 s. After adding the later-combination fault-injection regression, all 11 release benchmarks pass in both environments. Test ownership: 146 classes in nine groups. Offline frame viewport QA and strict documentation/link/browser checks pass, including the new native screenshot; the site contains 55 HTML pages. Installed smoke checks pass all 27 examples through save/open/analyze/export plus authenticated MCP discovery/edit/undo. Final wheel/source audits pass and all packaged runtime bytes match the tested independent installation. Human platform gates remain open.

Post-RC4 Plate Inspection Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-29. The published RC4 assets and plate input/file schemas remain unchanged; this does not publish RC5.

• Add a modeless separate-plate inspector with element-centre table, explicit element/natural-coordinate probes, material-face picking and linked outline/ cross marker. Opening holes and pan/zoom gestures do not change selection. Read signed Mx/My/Mxy and corrected Rect Qx/Qy with units, global coordinates, local axes and combination/snapshot identity. Quad shear remains unavailable; no general membrane or combined face-stress analysis is enabled.
• Accepted worker snapshots retain compact local recovery data, rather than a live solver. Probes use the corrected Rect polynomial or the engine's exact within-element bilinear Quad moment field without another solve. Compare off-centre/corner probes against explicit polynomial derivatives and actual solved elements, including non-square XY/XZ/YZ plates, origins, openings, signed/zero combinations and process serialization. Corner peaks remain unverified.
• Export all centres or one probe as CSV with xi/eta, local/global coordinates, units, axes, family/recovery description, case factors, snapshot/time/signature. Combination and display-unit changes retain the selection and snapshot; engineering edits/re-analysis clear inspection and disable stale exports.
• Focused surfaces: 116 tests, 26.094 s. Full locked suite with MCP: 925 tests, no skips, 243.636 s. The independently installed development wheel with PyNite 3.2 / Qt 6.12 passes 925 tests, no skips, 270.692 s. Test ownership: 144 classes in 9 groups.
• Automated native macOS checks use the independently installed wheel and actual worker: face picking, off-centre inspection, selected combination, both CSVs, unit retention and stale-result clearing pass. The guide includes an inspected native screenshot. Strict documentation/link checks cover 55 HTML pages; browser QA covers the new image, search, subpath and dark mode. Human platform gates remain separate. This batch changes no Three.js viewport assets.

Post-RC4 Coupled Validation Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-28. This source checkpoint does not create an RC5 tag or replace the published RC4 assets.

• Solve applied-traction membrane patches against independent plane-stress displacement/stress formulas, rather than prescribing the solution. Cover a non-square 120 × 80 in rectangle in XY/XZ/YZ, three mesh sizes, signed/zero combinations, uniaxial Poisson contraction and pure shear. Numerical errors remain below the explicit 1e-10 in / 1e-9 ksi regression bounds.
• Verify load transfer in both directions between a coincident beam and a nu=0 slender slab strip: pressure on the slab and equivalent member line loading. Independent narrow-strip EI-sum references cover three stiffness ratios and refinement. Fine-mesh maxima: 0.0781% displacement-field error relative to reference peak, 0.2587% beam midspan moment error, 0.1338% slab Mx L2 error. These are bounded flat-strip checks, not general diaphragm/composite or eccentric-connection validation. PyNite 3.0 and 3.2 reproduce these values.
• Quantify the lateral weak-drilling limitation independently: external MY residuals of about 0.296 / 0.604 / 1.046 kip-in on 12/48/192 slab elements agree with summed weak local drilling torque. All three actual application solves reject the result using unchanged equilibrium limits. No stiffness substitution or tolerance relaxation is introduced; Quad shear stays withheld.
• Correct coupled YZ outer-edge support and contour/probe coordinates to the upstream width-Z / height-Y convention, sharing the mapping with rectangle placement. Non-square, notch, remesh and off-centre native probe regressions prevent the square-fixture blind spot. No file-format change; review YZ edge assignments saved by earlier development builds.
• Focused surfaces: 110 tests, 21.360 s. Full locked suite with MCP: 919 tests, no skips, 237.281 s. An independently installed development wheel with PyNite 3.2 / Qt 6.12 also passes 919 tests, no skips, 262.552 s. Test ownership: 143 classes in 9 groups.
• Automated native macOS checks use the installed development wheel, actual analysis worker, corrected YZ support preview/plot and the retained All tools hidden-Results behaviour. They do not replace human OS-dialog/platform gates. Strict docs/link checking covers 55 HTML pages; documentation browser and offline mesh viewport checks pass. The guide includes a reproducible numeric report command and inspected scientific convergence figure.

Post-RC4 Plate Loading Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-28. No new release tag or registry upload is created by this checkpoint; the published RC4 files remain immutable.

• The separate rectangular plate tool supports up to 32 uniform pressure cases and 32 combinations. Its original primary-case Plate combination remains the baseline; named combinations mix cases. A staged editor handles pressures, factor maps, reference-preserving case renames and horizontal self-weight.
• Opt-in self-weight uses material weight density and global -Y. XZ plates carry the complete gravity as normal pressure, factored through its selected case. XY/YZ membrane gravity is rejected. Openings remove both pressure and weight. This does not enable general membrane, curved shells or a nonlinear solver.
• Every combination must have finite results and pass six-component global force/moment equilibrium before any snapshot is accepted. Cached selection, unit-only changes and their undo/redo preserve snapshot identity; engineering edits clear results. CSV and plot titles identify the selected combination and snapshot. Quad shear remains withheld; corrected Rect shear is separate.
• Plate files advance to version 3 with strict loading fields. Version 1/2 files migrate without adding cases or gravity. Frame/spatial file formats are unchanged.
• Independent checks cover signed/zero-result superposition for both element families in every plane, self-weight/Navier deflection within 1% on the chosen fine mesh, off-centre opening force/first-moment balance, migration and whole- snapshot rejection after force, moment or nonfinite errors. Native checks cover the worker, four combinations, staged controls, unit retention and CSV export.
• Full locked suite: 912 tests, no skips, 220.145 s. Installed development wheel with PyNite 3.2 / Qt 6.12: 912 tests, no skips, 242.787 s. Focused surfaces: 103 tests passed. Test ownership: 142 classes in 9 groups. Strict documentation build/link checks cover 55 HTML pages; browser QA verifies new screenshots, navigation, search, GitHub Pages subpath and dark mode.
• Native screenshots document the actual plate controls and selected-combination moment plot. Development wheel/source metadata and contents are audited. Stable 1.0 platform/manual feedback gates remain open.

Post-RC4 Coupled Slab Inspection Checkpoint¶

Development 1.0.0rc5.dev0, checked 2026-10-28; no RC5 tag or PyPI upload is created by this checkpoint. The RC4 assets remain unchanged.

• The existing Results dock now has local rectangular Quad moment/membrane-stress contours, element/natural-coordinate probes, axes, combination/unit/snapshot labels, provenance CSV and paginated MCP surface-element reads.
• Geometry-owned outer-edge global rigid restraints have presets/custom DOFs, staged preview, shared-joint unions, spring-conflict rejection, remesh/lifecycle handling, undo, viewport/report display and MCP editing. Spatial version 28 migrates old files with empty edge assignments; frame/slab coupling remains opt-in.
• Manufactured constant-strain/curvature recovery checks cover all three planes. An independently referenced thin square plate has fine interior moment errors below 0.6% and peak-deflection error below 0.01%; all meshes pass equilibrium. This bounded check does not validate general membrane/drilling response or corner peaks. The existing all-combination equilibrium gate remains strict.
• Raw Quad Qx/Qy has about 59% relative L2 error on the fine Navier check in both PyNite 3.0 and 3.2. Do not publish those numbers: coupled outputs omit them, plate-pilot Quad shear choices are disabled, and its shears payload is None. Rect shear recovery is separate. Reproduce with scripts/benchmark_coupled_surfaces.py --output /tmp/coupled-plate.json.
• Full locked Python suite: 902 tests, no skips, 219.723 s. Installed development wheel with PyNite 3.2 / Qt 6.12: 902 tests, no skips, 245.163 s. Final UI checks: 184 tests; focused recovery-error clearing and export re-enabling: 5 tests. Test ownership: 141 classes in 9 groups.
• Native macOS checks verified the slab tab, table/contour/probe, global edge presets and staged preview in the real desktop window. Light/dark result plots and the new documentation screenshots were visually inspected. Offline surface viewport QA includes support visibility/pixels, mesh families, picking/box selection and bounded resources. Strict docs: 55 HTML pages; browser QA passed screenshots, search, navigation, GitHub Pages subpath and dark mode.
• Wheel/source contents and metadata are audited. Platform/manual prerelease feedback gates remain open; this is not a stable 1.0 certification.

Post-RC4 Plate Recovery Validation¶

Unreleased 1.0.0rc5.dev0, verified 2026-10-28. The published RC4 files and tags are unchanged; the known limitation above applies to that release.

• 892 tests passed without skips in the locked environment (202.900 seconds, PyNiteFEA 3.0.0 / Qt 6.11.2) and the isolated development wheel (228.243 seconds, PyNiteFEA 3.2.0 / Qt 6.12.0), including real MCP wire tests.
• The affected surface group passed 85 tests, including the Rect subprocess solve and signed SI contour-array checks. New engineering tests cover intact Navier equivalence, independent Ritz basis convergence, opening deflection and patch moments, polynomial/rigid fields, mixed derivatives, cantilever strips, mesh refinement and plane/origin/factor invariance. Rect bending modifiers are rejected while their recipe/preview compatibility is retained.
• In the documented 512-element opening case, opening-edge deflection errors against the 13×13-mode reference are 0.214% Quad / 0.085% Rect. Largest material-patch mean moment errors are 0.636% / 0.398% respectively. The 11→13-mode reference patch-moment change is at most 0.158%. These are observed bounded-case comparisons, not universal element accuracy guarantees.
• scripts/benchmark_plate_recovery.py produces the numerical JSON record and standalone figure. Strict docs/link checks cover the new illustration and scope limits; the source/wheel audit includes the new recovery module, tests and reproducible benchmark script.

RC4 Release Validation¶

First Brecon release, 1.0.0rc4, prepared 2026-10-28 on macOS 27.2 arm64 / Python 3.12.13. Publication is explicitly authorized; this is a prerelease, with the stable-1.0 platform and independent-feedback gates still open.

• Full locked suite: 882 tests passed, no skips (200.177 seconds), including real MCP wire tests. Qt 6.11.2, PyNiteFEA 3.0.0, Matplotlib 3.11.1, NumPy 2.5.3, SciPy 1.18.1, MCP 2.3.0.
• Full installed-wheel suite: 882 tests passed, no skips (228.083 seconds). Qt 6.12.0, PyNiteFEA 3.2.0, Matplotlib 3.11.2, NumPy 2.5.4, SciPy 1.18.1, MCP 2.3.0. Final runtime files match the fully tested wheel byte for byte.
• Fresh base/MCP profiles verify the final wheel, both launch commands and all 26 examples through save/open/analyze/export and native-window startup. MCP verifies Brecon identity/version, authenticated discovery, atomic edit, undo and server shutdown. Isolated uv-tool installation checks both commands.
• Both viewport suites pass against installed Brecon assets: spatial frame editing/results/probes and all seven mesh families, mixed/surface-only scenes, picking, context recovery and resource limits. Headless Chromium uses SwiftShader; hardware evidence comes from the native check below.
• The visible installed RC4 Cocoa run recorded in the rename validation passed all workflows with ANGLE Metal / Apple M4 Max; its runtime is unchanged in the release build. Human OS-dialog confirmation remains the earlier RC3 evidence, not a new RC4 manual certification. Linux/Windows RC4 workflows are unverified.
• Strict docs/local links pass across 55 HTML pages; documentation browser QA covers images, search, existing Pages subpath, dark mode and layout. Source audit includes the publishing workflow and executable Brecon-only gate tests.
• The release carries a wheel, source archive and SHA256SUMS. GitHub and PyPI receive the same audited files; the workflow verifies tag/version, Brecon metadata and hashes before the environment-gated upload, without rebuilding.
• Publication completed through workflow 40713196939 after the owner's required pypi environment review. Public PyPI metadata and both artifact hashes match the reviewed build. A fresh installation from https://pypi.org/simple verifies the canonical and compatibility commands. This post-publication record does not move the sealed tag or replace its release artifacts.

Brecon Rename Validation¶

Development 1.0.0rc4, 2026-10-28. This is a source/package rename, not a new published release. Existing RC1/RC2/RC3 tags and assets remain unchanged.

• 258 focused tests passed without skips: UI, MCP, recovery compatibility, release metadata, documentation and tooling. A regression checks that the Brecon identity uses the original recovery path; version checks cover both entry modules, and real wire checks read both MCP reference URIs.
• Fresh independent base/MCP wheels passed all 26 examples through save/open/analyze/export, native-window startup, canonical brecon and legacy pynitegui version commands. MCP identifies itself as Brecon and passes authenticated discovery, atomic edit, undo and shutdown.
• An isolated local-wheel uv-tool installation creates both commands, each reporting Brecon 1.0.0rc4, without changing user tools/PATH.
• Visible installed-wheel Cocoa checks passed previews/generators/arrays, undo/redo, save/reopen with supplied paths, background solves, station queries, CSV/PDF/report preview, both official spatial frames and MCP lock/unlock. Runtime: macOS 27.2 arm64, Python 3.12.13, Qt 6.12.0, ANGLE Metal / Apple M4 Max. These automated checks do not replace human OS-dialog/platform verification; the previous RC3 owner confirmation below remains historical evidence.
• Strict documentation/link checks passed across 55 HTML pages. Browser QA checks the new SVG wordmark, refreshed About/Help screenshots, search, retained GitHub Pages subpath, dark mode and desktop layout. Its stale pre-coupled-slab search assertion now checks the current page title and experimental scope.
• Wheel/source audit checks both Python namespaces, both console commands, metadata, runtime assets, sources, citations and the new wordmark, with hashes. The initial rename validated both tag identities; the later Brecon-only publishing gate accepts brecon and rejects old pynitegui and mismatched tags in executable regression tests. No remote publishing was performed.

RC3 Validation Record¶

RC3 adds named groups/levels, scoped envelope summaries, continuous-segment governing-location queries, previewed transforms/generators/ node merging, and independent material G with two official spatial frames. New project saves use formats 25/26; older files migrate without changing their physics, and older applications reject new formats. Surface/plate recipes remain derived-G only. Stable 1.0 platform/feedback gates remain separate.

Verified on macOS 27.2 arm64 / Python 3.12.13, 2026-10-27–28:

• Locked environment: 879 tests passed, no skips (192.827 seconds), including real MCP wire checks. Qt 6.11.2, PyNiteFEA 3.0.0, Matplotlib 3.11.1, NumPy 2.5.3, SciPy 1.18.1, MCP 2.3.0.
• Fresh non-editable MCP wheel: 879 tests passed, no skips (192.374 seconds), with Qt 6.12.0, PyNiteFEA 3.2.0, Matplotlib 3.11.2, NumPy 2.5.4, SciPy 1.18.1 and MCP 2.3.0. These are observed dependency versions, not a guarantee for future releases.
• Separate base/MCP profiles passed all 26 examples through save/open, analysis and CSV/report export. MCP also passed authenticated SDK discovery, atomic edit, undo and server shutdown. Final artifacts are audited against the tagged source, including the repeatable desktop smoke script.
• An isolated local-wheel uv tool install passed; the generated entry point reports 1.0.0rc3. User tools and PATH were not changed.
• Both offline viewport suites passed against installed-wheel assets, covering frame results/probes and all seven mesh families, mixed scenes, picking, context recovery and resource limits. Chromium used SwiftShader; this is separate from the native hardware check below.
• The visible installed-wheel Cocoa check passed generator/array previews and apply, undo/redo, save/reopen with supplied paths, worker analysis, governing stations, CSV/PDF and native report preview, both official spatial frames, deformation/torsion and MCP lock/unlock/shutdown on an ephemeral port. Diagnostics report WebGL 2.0, ANGLE Metal, Apple M4 Max, context not lost. scripts/smoke_desktop.py records screenshots, versions and renderer details.
• Native validation found an analysis teardown race: the finished signal can precede deferred Python worker cleanup; deleting its QThread then blocked the GUI while that cleanup waited for the GIL. Frame and plate windows now poll a nonblocking join before deletion/control handover. A delayed-cleanup regression covers both workspaces; the complete native workflow passes with the fix.
• Strict documentation build and local-link/image/anchor checks passed across 55 HTML pages; documentation/tooling/release checks also pass. Pages was switched to Actions: run 40602180800 successfully deployed commit 953dc9a, with live content verified for modelling tools, examples and current formats.

The native automated check used visible QtTest/application commands and supplied file paths. After the desktop-control attachment timed out, the owner completed the remaining human checks in that installed RC3 window on 2026-10-28: Save As/Open through the real pickers, F5 analysis and the interactive Print Results dialog all worked. This confirmation arrived after the RC3 tag was created and is recorded separately without moving the tag or replacing assets. Linux and Windows visible installed-RC3 workflows are unverified. Those and independent prerelease feedback remain stable-1.0 gates. The consolidation creates an unpublished draft with wheel, source archive and SHA256SUMS; it does not publish RC3 to GitHub or PyPI.

RC2 Validation Record¶

Automated preparation verified on macOS 27.2 arm64 / Python 3.12.13, 2026-10-27:

• Locked editable environment: 831 tests passed, no skips, including real MCP wire tests (175.986 seconds). Qt 6.11.2, Matplotlib 3.11.1, PyNiteFEA 3.0.0, NumPy 2.5.3, SciPy 1.18.1, MCP 2.3.0.
• Non-editable wheel in a fresh MCP environment: 831 tests passed, no skips (202.407 seconds). Qt 6.12.0, Matplotlib 3.11.2, PyNiteFEA 3.2.0, NumPy 2.5.4, SciPy 1.18.1, MCP 2.3.0. These are observed dependency versions, not a promise about future releases.
• Base and MCP wheel profiles independently passed installed version/entrypoint, bundled assets, all 24 examples save/open/analyze/export, and offscreen native-window checks. MCP also passed authenticated SDK discovery, a real atomic edit, undo and server shutdown.
• A real uv tool install of the local wheel passed with isolated temporary tool/bin directories. The generated pynitegui --version reports 1.0.0rc2. User tool installations and PATH were not changed.
• Both offline viewport suites passed against installed-wheel assets: spatial frames/results/probes and all seven mesh families, mixed/surface-only scenes, picking, context recovery and bounded resources. Chromium used SwiftShader; these checks do not certify native Qt/WebEngine hardware rendering.
• Strict documentation build and local-link/image/anchor validation passed across 54 HTML pages. Distribution audit verifies matching runtime assets, metadata, licenses, citation files and source contents, with SHA256SUMS.
• The first full run found an obsolete two-tab assertion in the spatial results dialog test. It now explicitly checks Member Detail, Compare Combinations, Storey Drift and Combination Envelopes. The focused six tests and both full reruns pass; no product behaviour was changed to satisfy the assertion.

Visible macOS inspection was attempted, but the UI automation attachment to the installed Python window timed out. No visible-workflow gate is marked complete. Visible installed-app checks on the claimed platforms and independent prerelease feedback remain required before stable 1.0. Windows is currently unverified. The automated validation above performed no tag/push, release publication, PyPI upload or DOI deposit. Publication is a separate step from these checks.

Historical RC1 Validation Record¶

Automated preparation verified on Linux x86_64 / Python 3.12.15, 2026-10-09:

• Locked editable environment: 628 tests passed, no skips, including real MCP wire tests (160.753 seconds). Qt 6.11.2, Matplotlib 3.11.1, PyNiteFEA 3.0.0, MCP 2.3.0.
• Non-editable wheel, fresh dependency resolution outside the checkout: 628 tests passed, no skips (198.200 seconds). Qt 6.12.0, Matplotlib 3.11.2, PyNiteFEA 3.2.0, MCP 2.3.0. These are recorded versions, not a guarantee for untested future dependency releases.
• Base and MCP wheel profiles separately passed installed version/entrypoint, bundled asset, 13-example save/open/analyze/export and offscreen native-window checks. The MCP profile also passed authenticated SDK discovery, version identity, a schema-valid atomic edit, undo and server shutdown.
• The README's uv tool install command was verified using temporary tool directories; pynitegui --version reports 1.0.0rc1 without loading Qt.
• Installed wheel viewport assets passed Chromium 151 / SwiftShader desktop pixels, framing, picking, gizmo, drawing/dragging, supports, overlays, PNG, dark-mode and context-loss checks. This is software-rendered browser QA, not visible native Qt/WebEngine hardware validation.
• Initial full-suite attempts were interrupted after accumulated inactive MCP application filters caused severe cross-window/test overhead. The filter is now installed only during exclusive control mode and removed on unlock/stop (commit 7ad23f2); its new lifecycle regression and both complete reruns pass.

Earlier Linux viewport/Wayland checks and the user-supplied macOS M4 Max smoke test are useful history, not installed-RC certification. Visible installed-RC Linux/macOS workflows still require manual verification. Windows and standalone installers remain unverified. No tag, push, GitHub release or PyPI upload was made.