Maintain coverage tutorials

The phase-4 lessons cover the six additional algorithm families, four discrete/mixed representations, three practical cases and three advanced workflows. Sources are examples/tutorials/capability_examples.py and examples/tutorials/advanced_workflows.py; tools/build_coverage_tutorials.py owns downloads, code/output blocks, figures, observations and evidence under docs/assets/coverage/.

Regenerate from executed code

Use a clean non-editable environment with the exact coverage requirements, then run from the checkout:

OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMBA_NUM_THREADS=1 python tools/build_coverage_tutorials.py --write --work-dir ../coverage-recording-new
python tools/build_coverage_tutorials.py --check

The work directory must be new. It preserves complete ablation studies and the tuner’s original files. The checked-in observation JSON is a bounded derived dataset, not a canonical replay artifact. Source/asset hashes, package versions and base commit are recorded. Standard-library --check detects stale sources, downloads, generated blocks, dependency recipes and asset bytes; it does not rerun optimization or establish scientific validity.

Before creating any output, --write checks the resolved VAMOS module against the installed distribution and rejects source-checkout imports and editable distribution metadata. A matching version string alone does not establish that the published wheel supplied the code. --check remains independent of an installed VAMOS package.

The generated timing figure intentionally records real elapsed time, so a regeneration changes those values. Capture thread settings and platform metadata, inspect the output, and avoid interpreting incidental timing changes as regressions. Figures are generated with Matplotlib from real arrays or samples. Do not hand-edit outputs or asset hashes. Keep ordinary assets under 250 KB and the observation file under 750 KB.

Validate behavior and interpretation

Follow the repository’s validation tiers and documentation delivery. Run tests/docs/test_coverage_tutorials.py, the docs smoke manifest and both strict documentation builds. Test downloads copied outside the repository in the clean installed-wheel environment. Check meaningful invariants: exact evaluation counts, encoding domains, row alignment, emitted/declared constraints, non-overwrite behavior, changed ablation component and independently recomputed HV. Timings must not be subject to a speed threshold in CI.

Visually inspect representative two/three-objective plots, practical cases and the advanced figures. Check navigation, every new download and existing routes/anchors in the PR preview. Public examples use only curated facades; the tuner uses its Experimental CLI. Never label that CLI Stable merely because the wrapper is tested.

Coverage map

Capability	Precise reference	Verified teaching example
Nine built-in algorithms	Catalogue and capability matrix	Each catalogue row links to its complete guide and executable script.
Real decisions	Problem protocol	T01
Binary, integer, permutation, mixed decisions	Problem catalogue	Four representation lessons
Mixed constrained design	Constraint convention	Welded beam
Feature selection and model hyperparameters	Problem catalogue	Masks, mixed SVC settings
Algorithm configuration tuning	Experimental tuning	Train/validation/test seeds
Component ablation	StudySpec	Mutation removal
Backend performance	Backend roles	First and subsequent calls

This is teaching coverage of explicitly identified cases, not proof of the full Cartesian product of capabilities, convergence, model generalization or universal speedups.

The expanded menu uses Material's navigation.prune to avoid embedding inactive branches in every page. Inactive groups link to their first page; their complete contents appear after entering the group. This keeps the modular API hub within its existing HTML-size gate. Keep the first page of each group a useful entry point and verify navigation in both documentation engines.