Add R-style performance table with Great Tables and Reactable renderers

[uriahf]

What changed

• adds create_performance_table(), create_performance_table_times(), and render_performance_table() to the public Python API
• keeps the existing Polars prepare_performance_data() / prepare_performance_data_times() pipelines as the numerical source of truth
• adds Great Tables as the default renderer for Marimo and HTML output, styled to follow the original R Reactable table as closely as practical
• preserves Reactable as renderer="reactable" for Quarto/Jupyter contexts, including sortable columns, in-cell data bars, model badges, grouped performance columns, and expandable confusion matrices
• supports probability-threshold and PPCR stratification for the standard table
• supports time-dependent tables across one or multiple fixed time horizons and one or multiple censoring/competing-event heuristic sets
• keeps the time horizon, censoring heuristic, and competing-event heuristic visible in time-dependent output rather than collapsing distinct evaluation scenarios
• normalizes integer observed times at the public _times table wrapper boundary, while relying on the shared time-dependent pipeline for integer/float fixed-horizon handling
• adds a Marimo PR-preview demo using the Great Tables renderer, including a two-horizon time-dependent example
• adds tests covering both renderers, prepared-data rendering, empty input, PPCR stratification, invalid renderer values, multiple time horizons, and multiple heuristic sets

API

# Marimo-friendly default
create_performance_table(probs, reals)

# Time-dependent performance table
create_performance_table_times(
    probs,
    reals,
    times,
    fixed_time_horizons=[365, 730],
)

# Reactable remains available for Quarto/Jupyter
create_performance_table(probs, reals, renderer="reactable")
create_performance_table_times(
    probs,
    reals,
    times,
    fixed_time_horizons=[365],
    renderer="reactable",
)

Renderer choice

Great Tables is the Marimo-safe default and is deliberately styled toward the original R performance table: model colors, grouped performance columns, compact layout, proportional metric bars, predicted-positive bars, and diverging net-benefit bars. Reactable remains an optional renderer for Quarto/Jupyter; this PR does not pursue Reactable support in Marimo.

Time-dependent design

create_performance_table_times() is intentionally a thin presentation wrapper around prepare_performance_data_times(). It does not introduce a second statistical implementation. The renderers display the time horizon and heuristic dimensions explicitly so results from different time-dependent evaluation scenarios remain distinguishable.

Validation

• tests cover Great Tables and Reactable time-dependent output, multiple horizons, multiple heuristic sets, integer observed times through the public table wrapper, and integer/float fixed horizons through the shared pipeline
• the PR preview includes a Great Tables time-dependent example with fixed horizons 5 and 10 for visual review

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-12 10:57 UTC