Add HIF/Python interoperability and reproducible ecosystem benchmarks

[bregydoc]

Summary

Hyper can exchange scientific datasets through HIF without silently losing metadata or unsupported incidence semantics. A Rust-backed typed Python package supports HIF import/export and launches the native viewer. This PR also adds ecosystem research and reproducible benchmark evidence to guide positioning and roadmap decisions.

• Validate against a bundled, pinned HIF schema offline; preserve complete documents, typed IDs, weights and metadata. Viewer conversion rejects unsupported semantics with field locations.
• Detect HIF through the shared loader and support explicit --format for initial loads and watched reloads, preserving native/legacy behavior.
• Add PyO3/maturin bindings, generated stubs, an asynchronous viewer handle with snapshot cleanup, XGI/Julia examples and platform wheel CI.
• Ship hypergraph-viz (import hyper_viz) core wheels and a standalone source archive that excludes Bevy. Add hypergraph-viz-viewer native wheels through the exact-version [viewer] extra; discover installed executables through RECORD metadata without requiring environment activation. Explicit overrides take priority, then the companion, then PATH.
• Exclude unused Bevy audio/gamepad backends and audit Linux binaries without external library grafting, avoiding a pinned maturin binary-repair relocation defect. Link the Windows C runtime statically and inspect PE imports to avoid a separate Visual C++ runtime installer.
• Publish the companion before the core using configured Trusted Publishers, only after all core/source/desktop integration checks pass. Ordinary PR and manual runs never publish.
• Add an opt-in Bevy benchmark feature, core CPU example, canonical dataset generation and adapters for scientific figures, browser rendering and desktop workflows.
• Record 445 measured attempts across 89 configurations on one M5 Pro Mac, comparing Hyper with HyperNetX, XGI, Hypergraphx, 3d-force-graph, Cytoscape.js, Sigma, HyperGodot and Graphia. Include raw evidence, dataset/source checksums, failure counts, comparison tables and CSV, plus archived report regeneration.
• Document ecosystem findings in docs/ecosystem-comparison.md and the benchmark archive. Add docs/NOTEBOOK_ROADMAP.md: a future anywidget Canvas2D transport/selection prototype using a 7-node, 4-edge fixture, followed by a separate Bevy/WASM feasibility probe. No notebook API ships here.

Validation

• 71 core Rust tests and 52 Bevy library tests pass, along with all-target Clippy, formatting, generated-stub drift checks and strict Python typing.
• All 17 tests pass against locally installed 0.1.2 core and desktop wheels; executable discovery works with empty PATH. Strict PyPI metadata and wheel-content checks pass.
• Local macOS arm64 graphics capture, Python launch/close and temporary snapshot cleanup pass. The first desktop matrix passed all four native binary builds and installed-extra checks, including Linux software Vulkan/Xvfb graphics capture. Windows and macOS Intel checks cover CLI startup and discovery, without claiming GUI coverage.
• Linux wheel ELF inspection found only standard system runtime dependencies, with no ALSA/udev or grafted libraries. The previous Windows artifact exposed a Visual C++ runtime dependency; the new PE audit rejects it and gates the static-runtime rebuild.
• Core CI verifies clean CPython 3.10/3.14 installs and builds a wheel from the standalone source archive. The corrected 0.1.2 tag workflow passed every build and integration gate, then both Trusted Publishing uploads. A fresh public PyPI install verified typed IDs, metadata, load/save, exceptions, typing, empty-PATH discovery, and real macOS render/launch/close/cleanup. The final Windows PE imports only system DLLs, with no external Visual C++ runtime.
• Benchmark work includes 8 harness tests, 24 independent scientific conversion checks, archived checksums and byte-for-byte report regeneration. The Julia example was not runtime-tested.

Measurement limits

The tracks perform different work: scientific rasterization, browser callback cadence, native application intervals and Graphia import/save are separate metrics. Display cadence was not locked; these results do not establish a universal speed ranking. Native runs retain 16 timeouts across 120 attempts, including one on a small fixture; their cause remains unresolved. A source audit identified a quadratic hub-status lookup for follow-up profiling. Incorrect early Sigma refresh and clamped-window HyperGodot measurements are excluded diagnostics. Human task accuracy, input-to-visible latency, GPU duration and cross-platform benchmark rendering remain unmeasured.

Release status

hypergraph-viz 0.1.0 is published on PyPI. The unpublished python-v0.1.1 release was canceled after inspecting its Windows runtime dependency; that tag remains unchanged. Both hypergraph-viz 0.1.2 and hypergraph-viz-viewer 0.1.2 are published. The release workflow passed all gates and uploads; PyPI lists four wheels for each package and a core source archive. Its desktop installation command is python -m pip install --upgrade "hypergraph-viz[viewer]"; there are no downloads during show().

Related: ATO-2428. See docs/HIF.md, docs/PYTHON_RELEASE.md and scripts/benchmarks/README.md for integration, releases and reruns.