idleon-automation

Cleanup backlog (historical)

Open follow-ups have moved to GitHub issues — see label:area:cleanup. This file keeps the settled experiments and their conclusions so the “why” survives in repo history.

Hoops: click-position machinery

Conclusion from experiments (2026-05-03): click position has no measurable effect on ball trajectory.

• click_y sweep (170 → 490, 320px range) → ball_x_at_rim variation: 24px (within the ~20px noise floor at fixed click).
• click_extreme (window corners + center vs at-hoop) — ran the experiment and didn’t see meaningful trajectory differences either.

The earlier “click matters” reading was confounded by perturbation-sweep luck and a small number of flukey makes. The clean isolation experiments disproved it.

Removed (2026-05-03)

• CLICK_STRATEGY constants and all branches in _pick_click_position
• VARIED_CLICK_POSITIONS, CLICK_SWEEP_Y_OFFSETS, CLICK_EXTREME_POSITIONS constants
• EXPERIMENT_MODE env-var dispatch in minigames/hoops/main.py; removed the two click experiments from the hoops launcher entry (kept the general experiments slot for future use)
• tests/test_click_strategy.py deleted
• _pick_click_position inlined at the call site (was a one-liner)
• Launcher’s experiments config slot removed (was empty, plus its extra_env plumbing in _start_bot / _spawn)
• Stopped logging click_x / click_y to shots.db. Removed from schema, migration, dump_shots query, and tests. Old DBs still have the columns sitting there harmlessly (no DROP COLUMN migration — they take ~16 bytes per row, not worth the churn).
• CLAUDE.md note — added a “Hoops: established findings” section under Architecture covering click position, OCR + trajectory cross- check, and pre-shot OCR flakiness.

Kept

• click_x / click_y columns in shots.db — still useful per-shot diagnostic info even if we don’t vary them.
• Launcher’s general experiments config slot — useful framework for future tests (just not these specific ones).

Refactor: split predictors into their own module (done 2026-05-03)

• New common/predictor.py holds Predictor protocol + KnnPredictor + BivariatePredictor + fit_knn + fit_bivariate
• common/shot_log.py shrunk: removed KnnPredictor and fit_target_predictor. Added fetch_makes(conn, dir) which returns the rows; callers pick a predictor and fit themselves.
• minigames/hoops/main.py has a PREDICTOR_KIND constant ("knn" or "bivariate"); call site picks the right fit_*.
• predictor_kind column added to shots.db so per-shot history records which predictor produced the target_y.
• Tests split: tests/test_predictor.py for the new module, tests/test_shot_log.py slimmed to just logging concerns.