[Swissystem7]

Adds basketball/ (page) and basketball.js (engine) inside ScoutAI.

What it does

• Data: reads a user CSV (licence checkbox) or a synthetic example. Handles Basketball-Reference and Hebrew headers, 32:15 clock minutes, per-game files (turned into totals via G), TOT/2TM traded players counted once, and reports summary rows and repeated headers. A Round / מחזור column makes the file a game log (one row per player per game), summed to a season.
• Metric: transparent per-40 recipe with four components: defense (steals, blocks), rebounding (60% ORB, 40% DRB), playmaking (AST, AST/TO), scoring (points, TS%). Defaults 24/20/24/32 and a 250-minute floor, each tied to a cited source (TS% and Four Factors from Basketball-Reference, BPM 2.0, Franks et al. 2016, Medvedovsky 2020).
• Validation: team split (half the teams train, half test); always labelled "same season".
• "New line" test: any extra numeric column (body-language coding, fWHR from photos, touches) or PanimGuide is checked on the test teams for raw Spearman and partial Spearman beyond the metric, each against 1,000 shuffles. PanimGuide also gets a spelling probe.
• Start of season vs rest of season (part 6): on a game log, the metric is computed only on rounds up to a split and the target only on the rounds after it. The metric is compared with every simple early number (the target itself included), each against 1,000 shuffles. Typing a player's name (e.g. the MVP) shows his early rank by each measure and where he finished.
• Research section: face shape (Sato 2021, Welker 2015, Fujii 2016, Tsujimura 2013, Geniole 2015, Kosinski 2017) and body language (Kraus 2010, Moll 2010, Furley 2012/2014, Furley & Roth 2021, Lian 2026), plus a measurement protocol.
• Synthetic examples from tools/make-basketball-example.js (seeded): a season file with a pure-noise column and a planted positive-control column, and a game log (80 players, 14 rounds).
• Every page links to the basketball lab; the site-wide tests cover the new page.

Why rebounding was added (c63fcff)

A local check on 904 Israeli Premier League regular-season games (2021-22 to 2025-26, box scores from the league site, not committed) asked: after 8 rounds, where did the eventual MVP rank? The first recipe had no offensive rebounds and ranked centres far too low (Zach Hankins, MVP 2022-23: #45; the league's PIR had him #6). Rebounding gets its Four Factors share (20%); the other weights keep the old 30/30/40 proportions. It was not fitted to the data.

The result is mixed: the centres moved up (Onuaku 10 → 2, Hankins 45 → 13, Sorkin 7 → 4), but the guard Jared Harper moved down (10 → 26, 14 → 28). Early PIR still predicted rest-of-season PIR better than the metric in every season. Further recipe changes should be judged on seasons not yet looked at (2016-17 to 2020-21).

Results on the synthetic files

• New-line test: the random column stays inside the noise band; the planted column is caught raw but adds nothing beyond the metric. PanimGuide is noise, and its score changes with name spelling.
• Start-of-season test on the synthetic game log: the metric from rounds 1-4 predicts later PIR (ρ = 0.75) a little better than early PIR (0.68).

Testing

• npm test: 185 pass, 0 fail.
• CodeQL alert on a test regexp (bad tag filter) fixed in 7119869 and e3ce4c6.
• Checked in a browser at desktop and phone width: no console errors, no horizontal scroll; sliders, the split slider and the player field work.

🤖 Generated with Claude Code