"""Refinement-count sweep on a controller checkpoint, at several execution lengths.

Exists because ``run_ablations.py:67`` hardcodes ``data/runs/controller/controller.pt``
-- the *terminal-only* controller that ``progress.md`` records as superseded. Every
K-sweep row in ``data/runs/eval/results.jsonl`` is therefore measured on a checkpoint
with a known-broken objective, and the arrival+hold controller that replaced it
(``data/runs/ah_hold0.5``, 94% at rh=1) has never been swept at all.

The question this answers: does refinement count degrade success for a controller
that already does joint whole-plan refinement over a full ``rollout_plan``? If it
does, "causal one-step commitment" cannot be the cause of the recursive planner's
T-inversion, because this architecture has no causal commitment to begin with.

Each row goes through the unmodified ``eval_controller.py``, so it shares the
harness, the wrappers and -- at a fixed seed -- the held-out start/goal pairs with
every historical row. Results land in their own directory; the historical files are
not appended to.
"""

import argparse
import subprocess
import sys
from pathlib import Path

PY = sys.executable
ROOT = Path(__file__).resolve().parents[1]


def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument(
        '--controller', default='data/runs/ah_hold0.5/controller.pt'
    )
    p.add_argument('--refinements', type=int, nargs='+',
                   default=[0, 1, 2, 3, 5, 8])
    p.add_argument('--receding-horizon', type=int, nargs='+', default=[1, 5])
    p.add_argument('--seeds', type=int, nargs='+', default=[42, 43, 44])
    p.add_argument('--goal-offset', type=int, nargs='+', default=[25])
    p.add_argument('--num-eval', type=int, default=50)
    p.add_argument('--eval-budget', type=int, default=50)
    p.add_argument('--out', default='data/runs/eval_refine_sweep')
    return p.parse_args()


def main():
    args = parse_args()
    total = (
        len(args.refinements) * len(args.receding_horizon)
        * len(args.seeds) * len(args.goal_offset)
    )
    done = 0

    for q in args.goal_offset:
        for rh in args.receding_horizon:
            for seed in args.seeds:
                for k in args.refinements:
                    done += 1
                    cmd = [
                        PY, 'scripts/eval_controller.py',
                        '--controller', args.controller,
                        '--refinements', str(k),
                        '--receding-horizon', str(rh),
                        '--goal-offset', str(q),
                        '--seed', str(seed),
                        '--num-eval', str(args.num_eval),
                        '--eval-budget', str(args.eval_budget),
                        '--out', args.out,
                    ]
                    print(
                        f'\n[{done}/{total}] q={q} rh={rh} seed={seed} K={k}',
                        flush=True,
                    )
                    # one bad row should not discard the runs queued behind it
                    code = subprocess.run(cmd, cwd=ROOT).returncode
                    if code != 0:
                        print(f'!! exited {code}, continuing', flush=True)

    print(f'\nappended to {Path(args.out) / "results.jsonl"}')


if __name__ == '__main__':
    main()