"""Fresh v5: recover B0 evaluation without repeating its1000 updates, then B1."""
from pathlib import Path
import argparse
import json
import os
import subprocess
import sys

PROJECT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(PROJECT))
from pallet.mask_model_improvement_device_recovery import V4, V5


def prepare():
    import shutil
    from pallet.mask_model_training_data import verify_dataset, read, write, digest
    manifest,_=verify_dataset(V4)
    V5.mkdir(mode=0o700)
    names=set(manifest['files'])|{'dataset_manifest.json','private/preservation_before.json',
        'private/preservation_after_freeze.json','research_permit.json'}
    for name in sorted(names):
        path=V5/name;path.parent.mkdir(mode=0o700,parents=True,exist_ok=True)
        shutil.copyfile(V4/name,path);path.chmod(0o600)
        if digest(path)!=digest(V4/name):raise ValueError('Frozen copy differs')
    verify_dataset(V5)
    write(V5/'RECOVERY_SCOPE.json',dict(reason='After actual1000 updates, Lightning teardown moved model toCPU before final GPU inference',
        failed_B0='outputs/pallet_mask_model_improvement_v4/B0', failure_sha256=digest(V4/'B0/failure.json'),
        B0='recover fixed best native dev/cal evaluation only, zero new training updates',
        B1='original fixed matrix point-loss run, fresh generic seed42/1000updates; same model/matcher/EMA/optimizer/data',
        repair='process-local rf_predictions moves model to input CUDA device after trainer teardown; inference only',
        data_byte_identical_to_v3_v4=True, future_human_saves_included=False,
        old_code_or_receipts_overwritten=False, accepted_quality=None, production_activation=False))


def queue():
    from pallet.mask_model_training_data import write,GPU_UUIDS
    write(V5/'queue_m2_start.json',dict(pid=os.getpid(),runs=['B0_recovery','B1'],
        B0_new_optimizer_steps=0,B1_expected_optimizer_steps=1000,cuda_visible_devices=GPU_UUIDS['m2']))
    done=[]
    env=dict(os.environ,CUDA_VISIBLE_DEVICES=GPU_UUIDS['m2'],OMP_NUM_THREADS='4',MKL_NUM_THREADS='4',
        TMPDIR='/tmp/opencode',PYTHONUNBUFFERED='1')
    for name,command in [('B0_recovery','recover'),('B1','worker')]:
        with (V5/f'{name}.log').open('xb') as stream:
            process=subprocess.Popen([sys.executable,str(Path(__file__).resolve()),command],env=env,
                stdout=stream,stderr=subprocess.STDOUT)
            write(V5/f'{name}_process.json',dict(pid=process.pid,name=name,cuda_visible_devices=GPU_UUIDS['m2']))
            code=process.wait()
        done.append(dict(name=name,returncode=code));print(json.dumps(done[-1]),flush=True)
        if code:
            write(V5/'queue_m2_completion.json',dict(state='failed_or_partial_not_complete',runs=done))
            return code
    write(V5/'queue_m2_completion.json',dict(state='completed_recovery_and_remaining_B1',runs=done))
    return 0


if __name__=='__main__':
    parser=argparse.ArgumentParser(description=__doc__)
    parser.add_argument('command',choices=['prepare','declare','queue','recover','worker','wait-start'])
    args=parser.parse_args()
    if args.command=='prepare':prepare()
    elif args.command=='declare':
        from pallet.mask_model_improvement import predeclare
        predeclare(V5)
    elif args.command=='queue':raise SystemExit(queue())
    elif args.command=='recover':
        from pallet.mask_model_improvement_device_recovery import recover_b0
        recover_b0()
    elif args.command=='wait-start':
        from run_mask_model_improvement import wait_receipt
        wait_receipt(V5,['B1'])
    else:
        from pallet.mask_model_improvement import worker
        from pallet.mask_model_improvement_retry import config_compatibility
        from pallet.mask_model_improvement_device_recovery import postfit_device_compatibility
        with config_compatibility(),postfit_device_compatibility():worker('B1',V5)
