Expose PZ two-jet polynomial trace in benchmark notebook

[MoritzMaibaum]

Motivation

• Make the polynomial-zonotope two-jet propagation inspectable by exposing the actual propagated input/domain and per-layer Y, J, and H polynomials prior to squaring and integration so users can debug monomial growth and approximation behavior.
• Keep notebook output opt-in and size-bounded to avoid overwhelming interactive sessions as monomial counts grow.

Description

• Updated notebooks/affine_pz_twojet_adaquad_benchmarks.ipynb to call model.eval_pz_twojet(cell.domain, return_trace=True) and reuse trace_result.final for subsequent norm integrand construction so the traced polynomials correspond to the computed norms.
• Added imports for inline display and rendering (IPython.display.Markdown, display) and pz_to_latex, and added configurable display limits TRACE_MAX_TERMS, TRACE_PRECISION, and TRACE_COMPONENT_LIMIT plus TRACE_DISPLAY_MARKDOWN and TRACE_OUTPUT_PATH for a repo-local notebooks/generated/pz_twojet_trace.md output.
• Implemented helper functions that summarize term counts and max degree, render a size-limited set of scalar components with pz_to_latex, write the generated Markdown/LaTeX trace file, and optionally display it in the notebook; kept the existing y_sq, j_sq, h_sq computation after trace rendering.

Testing

• Ran git diff --check to ensure no whitespace/git-diff issues and it passed.
• Validated notebook JSON formatting with python -m json.tool notebooks/affine_pz_twojet_adaquad_benchmarks.ipynb which succeeded, and compiled each code cell via Python compile(...) to check for syntax errors which also succeeded.
• Attempted to execute the modified notebook cell(s) in this environment to exercise the trace writer, but a partial run that executes notebook setup cells failed due to missing runtime dependencies (numpy/matplotlib) so full runtime execution was not completed here.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6f1b5f43e9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Handle tuple-backed PZ components in the trace renderer

When running this notebook as written, domain = IntervalTensor.from_bounds(...) stores bounds as Python tuples, and PZIntegrationCell.from_affine_box(domain) builds a tuple-backed PolynomialZonotope; the first _render_pz_components("X", cell.domain) then reaches this line with z.center as a tuple, so .device/.shape raises AttributeError before the direct PZ norms are computed. Use z.shape and a fallback scalar-index iterator (as the diagnostics cell already does), or keep the integration-cell domain torch-backed.

Useful? React with 👍 / 👎.