v1.101.0

Added

• Collect/render split for out-of-process report rendering. Building a report allocates large holoviews/panel/bokeh object graphs; when CPython's cyclic GC traverses them alongside foreign live C-extension state (e.g. ROS 2 rclpy/DDS), the process can segfault. The split lets rendering happen in a clean process that never imported the foreign extension:
  • plot_sweep(auto_plot=...) — new parameter (defaults to None, deferring to run_cfg.auto_plot, itself True). When False, the sweep runs and regression detection is computed but no plotting objects are constructed.
  • Bench.collect(...) — thin wrapper for plot_sweep(auto_plot=False); returns a fully-populated, picklable BenchResult (dataset + regression report).
  • bencher.save_result() / load_result() / render_report() (new bencher.render module) — persist a collected result and render the HTML report from it, optionally in a separate process via python -m bencher.render <result> <out_dir>.
• Three test layers guarding the split against divergence from the normal plot_sweep path: parity tests (collect() computes the same dataset/regression as plot_sweep()), a breadth round-trip over every generated result type (save → load → render to HTML, plus a real-subprocess media test), and the BENCHER_FORCE_SPLIT_RENDER=1 switch that reroutes every auto-plot report build through serialize/render-from-loaded so pixi run test-split re-runs the whole suite over the split pipeline (own parallel py313-only ci-split job).

Changed

• BenchReport.append_result() gained an optional render_from= argument so a caller can register one result for identity-based tab routing while building the tab pane from another (used by the forced-split path).