PercentileBar always renders percentiles on the upward scale; inversion happens upstream

Context

Lower-is-better metrics (errors, fouls committed, tackles lost) can be rendered two ways:

• Invert inside the chart — errors-percentile-15 becomes a bar at 85%. The visual reads “good player”. But now the chart is doing a semantic flip that depends on a direction field; consumers who accidentally double-invert (flip upstream, flip again in the chart) get silent wrong answers.
• Render the raw percentile, badge the direction — errors-percentile-15 renders a short bar with a “lower is better” annotation. The visual reads “bottom-15% errors = great”. One direction, no double-flip risk.

Double-flip silently-wrong is the worse failure. Ambiguous bars with no badge are the weaker but more common failure mode.

Decision

• percentile is always display-scale (0 at the left, 100 at the right).
• originalDirection: "lower" renders an inversion badge and an accessibility note, never flips the bar.
• Callers (scouting layer, adapter) own any inversion needed before passing percentiles to the chart.
• aria-label reads “Player A, errors, 15th percentile (lower is better)” — screen readers get the direction without visual inference.

Consequences

• Callers pre-compute percentiles against the correct direction; the chart stops being a semantic trap.
• Inversion bugs become visible: a bar at 15% with no “lower is better” badge is obviously an un-inverted error metric, not a silent wrong answer.
• Future provider adapters that surface percentiles directly (some ship playerScout endpoints with pre-computed percentiles) feed the chart without transformation.
• Composition with RadarChart / PizzaChart follows the same rule — all percentile-consuming charts assume display-scale input.