Diagnose out-of-plane motion and document tutorial dimensionality

[marpaia]

jitter_z=False removes the newly added Z perturbation during division; inherited tilt, three-dimensional contact normals, soft walls and flow can still produce Z motion. Add a reproducible diagnostic that records initialization and each native stage, both center-Z and direction-Z changes, initial geometry/constraints, model provenance and the first departure with before/after state. Probe instrumentation restores native methods and preserves the random stream.

Seven shared-backend fixtures separate planar controls, crossing/coincident contacts, inherited tilt, division, wall relaxation and vertical flow. CPU and Metal crossing rods separate to Z≈±0.4 without division. Default soft-wall relaxation stops within its residual tolerance; tighter solves reduce penetration below 1e-6. A source audit documents every tutorial's dimensionality, and a nine-case CPU/Metal trajectory matrix distinguishes intentional XYZ jitter from bounded planar observations.

No solver defect was demonstrated. This implements the investigation's documented-diagnosis outcome and specifies a separate strict-2D feature covering translation, rotation, contacts, flow and checkpoint state. It does not silently redefine jitter or claim finite-height walls enforce exact planarity. The original older-version trajectory cannot be uniquely attributed without its inputs.

Validation: 180 focused tests passed on CPU and real Apple Metal; two explicit CUDA skips. Ruff, strict Python typing (one pre-existing native-source warning), diagnostic RNG/restoration regression and Markdown links passed. CUDA source was reviewed; NVIDIA execution remains unverified.

Originally reviewed against marpaia/22 so the tutorial audit includes corrected founder initialization. Prerequisites were integrated before retargeting to master.

Closes #14.

Independent review is complete against the original acceptance criteria and intended review base. The review report records per-criterion evidence, both corrected campaign findings, exact tested heads, landing order and platform limits.

Merged into master in the approved campaign order: #26, #27, #28, #31, #34, #36, #29, #30, #33, #35, #37, #32. The final tree preserves the reviewed implementation and integration regressions. Prerequisites were integrated with merge commits, and the frozen review bases were retained.

Landing validation: 534 Python tests passed (57 environment-dependent skips), 115 viewer tests passed, Ruff lint, strict type checks, viewer build/format checks, and actual Chromium feature/integration checks passed. Pyright retains the existing native-extension source warning. A Windows-only documentation link-test encoding failure was reproduced and fixed with an explicit UTF-8 read; the corrected final PR heads passed Windows CLI/shutdown checks and CUDA compilation. NVIDIA runtime execution remains unverified.