5.py

from __future__ import annotations
from dataclasses import dataclass
import json
import inspect
import math
import os
import random
import shutil
import time
import traceback
import gc
from contextlib import nullcontext
from pathlib import Path
from typing import Any

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import optuna
import pandas as pd
import segmentation_models_pytorch as smp
import torch
import torch.nn as nn
import torch.nn.functional as F
from optuna.storages import RDBStorage, RetryFailedTrialCallback
from PIL import Image as PILImage
from scipy import ndimage
from torch.optim import Adam, AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import DataLoader, Dataset
from tqdm.auto import tqdm

"""=============================================================================
EDIT ME
=============================================================================
"""

PROJECT_DIR = Path(__file__).resolve().parent #Path("/content/drive/MyDrive/SDP_ultrasound") ## 

RUNS_ROOT = PROJECT_DIR / "runs"
HARD_CODED_PARAM_DIR = PROJECT_DIR
MODEL_NAME = "EfficientNet_Strategy3_Refinement"

DATASET_NAME = "BUSI_with_classes"  # "BUSI" or "BUSI_with_classes"
SUPPORTED_DATASET_NAMES = ("BUSI", "BUSI_with_classes")
DATA_ROOT = PROJECT_DIR / DATASET_NAME
BUSI_WITH_CLASSES_SPLIT_POLICY = "stratified"  # "balanced_train" or "stratified"
SUPPORTED_BUSI_WITH_CLASSES_SPLIT_POLICIES = ("balanced_train", "stratified")

STRATEGIES = [3] #[1, 2, 3, 4, 5]
DATASET_PERCENTS = [0.1, ] #, 0.5, 1.0] #, 0.5, 1.0]
SPLIT_TYPE = "80_10_10"
SUPPORTED_SPLIT_TYPES = ("80_10_10", "70_10_20")
DATASET_SPLITS_JSON = PROJECT_DIR / "dataset_splits.json"
DATASET_SPLITS_VERSION = 1
NUM_TRIALS = 25
STUDY_DIRECTION = "maximize"
STUDY_OBJECTIVE_METRIC = "val_iou"
BEST_CHECKPOINT_METRIC = "val_iou"

SEED = 42
IMG_SIZE = 128
BATCH_SIZE = 2 # Recommended to prevent OOM
NUM_WORKERS = 0  # Recommended with RAM-preloaded datasets to avoid worker RAM duplication.
USE_PIN_MEMORY = True
USE_PERSISTENT_WORKERS = False
PRELOAD_TO_RAM = True

SMP_ENCODER_NAME = "efficientnet-b0"
SMP_ENCODER_WEIGHTS = "imagenet"
SMP_ENCODER_DEPTH = 5
SMP_ENCODER_PROJ_DIM = 128
SMP_DECODER_TYPE = "Unet"
BACKBONE_FAMILY = "smp"  # "smp" or "custom_vgg"
VGG_FEATURE_SCALES = 4
VGG_FEATURE_DILATION = 1

USE_IMAGENET_NORM = True
REPLACE_BN_WITH_GN = True
GN_NUM_GROUPS = 8
NUM_ACTIONS = 5

STRATEGY_1_MAX_EPOCHS = 100
STRATEGY_2_MAX_EPOCHS = 100
STRATEGY_3_MAX_EPOCHS = 100
STRATEGY_4_MAX_EPOCHS = 100
STRATEGY_5_MAX_EPOCHS = 100
VALIDATE_EVERY_N_EPOCHS = 5
CHECKPOINT_EVERY_N_EPOCHS = 10
SAVE_LATEST_EVERY_EPOCH = True
EARLY_STOPPING_PATIENCE = 0
VERBOSE_EPOCH_LOG = False

DEFAULT_HEAD_LR = 1e-4
DEFAULT_ENCODER_LR = 1e-5
DEFAULT_WEIGHT_DECAY = 1e-4
DEFAULT_TMAX = 5
DEFAULT_GAMMA = 0.95
DEFAULT_CRITIC_LOSS_WEIGHT = 0.5
DEFAULT_ENTROPY_ALPHA_INIT = 0.2
DEFAULT_ENTROPY_TARGET_RATIO = 0.1
DEFAULT_ENTROPY_LR = 3e-4
DEFAULT_CE_WEIGHT = 0.5
DEFAULT_DICE_WEIGHT = 0.5
DEFAULT_DROPOUT_P = 0.2
DEFAULT_GRAD_CLIP_NORM = 6.0
DEFAULT_MASK_UPDATE_STEP = 0.1
DEFAULT_REFINEMENT_DICE_REWARD_WEIGHT = 0.25
DEFAULT_REFINEMENT_PRIOR_REWARD_WEIGHT = 0.05
ATTENTION_MAX_TOKENS = 1024
ATTENTION_MIN_POOL_SIZE = 16

SCHEDULER_FACTOR = 0.5
SCHEDULER_PATIENCE = 5
SCHEDULER_THRESHOLD = 1e-3
SCHEDULER_MIN_LR = 1e-5

HEAD_LR_RANGE = (1e-5, 3e-3)
ENCODER_LR_RANGE = (1e-6, 3e-3)
WEIGHT_DECAY_RANGE = (1e-6, 1e-2)
TMAX_RANGE = (3, 10)
ENTROPY_LR_RANGE = (1e-5, 1e-3)
DROPOUT_P_RANGE = (0.0, 0.5)

USE_TRIAL_PRUNING = False
TRIAL_PRUNER_WARMUP_STEPS = 5
LOAD_EXISTING_STUDIES = True
SKIP_EXISTING_FINALS = False
RUN_OPTUNA = False
RESET_ALL_STUDIES_EACH_RUN = False
USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF = False

EXECUTION_MODE = "train_eval"  # "train_eval" or "eval_only"
EVAL_CHECKPOINT_MODE = "latest"  # "latest", "best", or "specific"
EVAL_SPECIFIC_CHECKPOINT = ""
STRATEGY2_CHECKPOINT_MODE = "specific"  # "latest", "best", or "specific"
STRATEGY2_SPECIFIC_CHECKPOINT = {
    0.1: "./runs/EfficientNet/pct_10/strategy_2/final/checkpoints/latest.pt",
    0.5: "./runs/EfficientNet/pct_50/strategy_2/final/checkpoints/latest.pt",
    1.0: "./runs/EfficientNet/pct_100/strategy_2/final/checkpoints/latest.pt"
}
STRATEGY3_BOOTSTRAP_FROM_STRATEGY2 = True
TRAIN_RESUME_MODE = "off"  # "off", "latest", "best", or "specific"
TRAIN_RESUME_SPECIFIC_CHECKPOINT = ""
OPTUNA_HEARTBEAT_INTERVAL = 60
OPTUNA_HEARTBEAT_GRACE_PERIOD = 180
OPTUNA_FAILED_TRIAL_MAX_RETRY = 3

USE_AMP = True
AMP_DTYPE = "bfloat16"  # "auto", "bfloat16", or "float16"
USE_CHANNELS_LAST = False
USE_TORCH_COMPILE = True
STEPWISE_BACKWARD = True
ALLOW_TF32 = True

RUN_SMOKE_TEST = True
SMOKE_TEST_SAMPLE_INDEX = 0
RUN_OVERFIT_TEST = False
OVERFIT_N_BATCHES = 2
OVERFIT_N_EPOCHS = 100
OVERFIT_HEAD_LR = 1e-3
OVERFIT_ENCODER_LR = 1e-4
OVERFIT_PRINT_EVERY = 5
CONTROLLED_MASK_THRESHOLD = 0.50

REQUIRED_HPARAM_KEYS = ("head_lr", "encoder_lr", "weight_decay", "dropout_p", "tmax", "entropy_lr")

"""=============================================================================
IF OPTUNA IS OFF --> USE ME
=============================================================================
"""

# Key format: "<strategy>:<dataset_percent_label>"
# Each value is a JSON filename in HARD_CODED_PARAM_DIR containing the required hyperparameters.
MANUAL_HPARAMS_IF_OPTUNA_OFF: dict[str, str] = {
    "3:10": "strat3_pct10.json",
}

"""=============================================================================
RUNTIME SETUP
=============================================================================
"""

torch.set_float32_matmul_precision("high")
if torch.cuda.is_available():
    torch.backends.cuda.matmul.allow_tf32 = ALLOW_TF32
    torch.backends.cudnn.allow_tf32 = ALLOW_TF32
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

def select_runtime_device() -> tuple[torch.device, str]:
    if torch.cuda.is_available():
        return torch.device("cuda"), "cuda"

    mps_backend = getattr(torch.backends, "mps", None)
    if mps_backend is not None and mps_backend.is_available():
        try:
            _probe = torch.zeros(1, device="mps")
            del _probe
            return torch.device("mps"), "mps"
        except Exception as exc:
            print(f"[Device] MPS detected but failed to initialize ({exc}). Falling back to CPU.")

    return torch.device("cpu"), "cpu"

DEVICE, DEVICE_FALLBACK_SOURCE = select_runtime_device()
CURRENT_JOB_PARAMS: dict[str, Any] = {}

@dataclass(frozen=True)
class RuntimeModelConfig:
    backbone_family: str
    smp_encoder_name: str
    smp_encoder_weights: str | None
    smp_encoder_depth: int
    smp_encoder_proj_dim: int
    smp_decoder_type: str
    vgg_feature_scales: int
    vgg_feature_dilation: int

    @classmethod
    def from_globals(cls) -> RuntimeModelConfig:
        return cls(
            backbone_family=str(BACKBONE_FAMILY).strip().lower(),
            smp_encoder_name=str(SMP_ENCODER_NAME),
            smp_encoder_weights=SMP_ENCODER_WEIGHTS,
            smp_encoder_depth=int(SMP_ENCODER_DEPTH),
            smp_encoder_proj_dim=int(SMP_ENCODER_PROJ_DIM),
            smp_decoder_type=str(SMP_DECODER_TYPE),
            vgg_feature_scales=int(VGG_FEATURE_SCALES),
            vgg_feature_dilation=int(VGG_FEATURE_DILATION),
        )

    @classmethod
    def from_payload(cls, payload: dict[str, Any] | None) -> RuntimeModelConfig:
        payload = payload or {}
        return cls(
            backbone_family=str(payload.get("backbone_family", "smp")).strip().lower(),
            smp_encoder_name=str(payload.get("smp_encoder_name", SMP_ENCODER_NAME)),
            smp_encoder_weights=payload.get("smp_encoder_weights", SMP_ENCODER_WEIGHTS),
            smp_encoder_depth=int(payload.get("smp_encoder_depth", SMP_ENCODER_DEPTH)),
            smp_encoder_proj_dim=int(payload.get("smp_encoder_proj_dim", SMP_ENCODER_PROJ_DIM)),
            smp_decoder_type=str(payload.get("smp_decoder_type", SMP_DECODER_TYPE)),
            vgg_feature_scales=int(payload.get("vgg_feature_scales", VGG_FEATURE_SCALES)),
            vgg_feature_dilation=int(payload.get("vgg_feature_dilation", VGG_FEATURE_DILATION)),
        )

    def validate(self) -> RuntimeModelConfig:
        if self.backbone_family not in {"smp", "custom_vgg"}:
            raise ValueError(f"BACKBONE_FAMILY must be 'smp' or 'custom_vgg', got {self.backbone_family!r}")
        if self.vgg_feature_scales not in {3, 4}:
            raise ValueError(f"VGG_FEATURE_SCALES must be 3 or 4, got {self.vgg_feature_scales}")
        if self.vgg_feature_dilation < 1:
            raise ValueError(f"VGG_FEATURE_DILATION must be >= 1, got {self.vgg_feature_dilation}")
        if self.smp_encoder_depth < 1:
            raise ValueError(f"SMP_ENCODER_DEPTH must be >= 1, got {self.smp_encoder_depth}")
        if self.smp_encoder_proj_dim < 0:
            raise ValueError(f"SMP_ENCODER_PROJ_DIM must be >= 0, got {self.smp_encoder_proj_dim}")
        return self

    def to_payload(self) -> dict[str, Any]:
        return {
            "backbone_family": self.backbone_family,
            "smp_encoder_name": self.smp_encoder_name,
            "smp_encoder_weights": self.smp_encoder_weights,
            "smp_encoder_depth": self.smp_encoder_depth,
            "smp_encoder_proj_dim": self.smp_encoder_proj_dim,
            "smp_decoder_type": self.smp_decoder_type,
            "vgg_feature_scales": self.vgg_feature_scales,
            "vgg_feature_dilation": self.vgg_feature_dilation,
        }

    def backbone_tag(self) -> str:
        return self.backbone_family

    def backbone_display_name(self) -> str:
        if self.backbone_family == "custom_vgg":
            return f"Custom VGG (scales={self.vgg_feature_scales}, dilation={self.vgg_feature_dilation})"
        return f"SMP {self.smp_encoder_name}"

def current_model_config() -> RuntimeModelConfig:
    return RuntimeModelConfig.from_globals().validate()

"""=============================================================================
UTILITIES
=============================================================================
"""

def banner(title: str) -> None:
    line = "=" * 80
    print(f"\n{line}\n{title}\n{line}")

def section(title: str) -> None:
    print(f"\n{'-' * 80}\n{title}\n{'-' * 80}")

def ensure_dir(path: str | Path) -> Path:
    path = Path(path).expanduser().resolve()
    path.mkdir(parents=True, exist_ok=True)
    return path

def save_json(path: str | Path, payload: Any) -> None:
    path = Path(path)
    ensure_dir(path.parent)
    with path.open("w", encoding="utf-8") as f:
        json.dump(payload, f, indent=2)

def load_json(path: str | Path) -> Any:
    with Path(path).open("r", encoding="utf-8") as f:
        return json.load(f)

def _format_history_log_value(key: str, value: Any) -> str:
    if isinstance(value, float):
        if key == "lr" or key.endswith("_lr"):
            return f"{value:.6e}"
        return json.dumps(value)
    return json.dumps(value)

def format_history_log_row(row: dict[str, Any]) -> str:
    return ", ".join(f"{key}={_format_history_log_value(key, value)}" for key, value in row.items())

def _format_epoch_metric(value: Any, *, scientific: bool = False) -> str:
    if value is None:
        return "null"
    if isinstance(value, (float, int, np.floating, np.integer)):
        value = float(value)
        return f"{value:.6e}" if scientific else f"{value:.4f}"
    return str(value)

def format_concise_epoch_log(row: dict[str, Any], *, best_val_iou: float) -> str:
    fields = (
        ("train_loss", row.get("train_loss"), False),
        ("train_iou", row.get("train_iou"), False),
        ("train_entropy", row.get("train_entropy"), False),
        ("val_loss", row.get("val_loss"), False),
        ("val_iou", row.get("val_iou"), False),
        ("val_dice", row.get("val_dice"), False),
        ("head_lr", row.get("lr"), True),
        ("encoder_lr", row.get("encoder_lr"), True),
        ("best_val_iou", best_val_iou, False),
    )
    return ", ".join(
        f"{name}={_format_epoch_metric(value, scientific=scientific)}"
        for name, value, scientific in fields
    )

def set_global_seed(seed: int = 42) -> None:
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

def stable_int_from_text(text: str) -> int:
    value = 0
    for byte in text.encode("utf-8"):
        value = (value * 131 + byte) % (2 ** 31 - 1)
    return value

def seed_worker(worker_id: int) -> None:
    del worker_id
    worker_seed = torch.initial_seed() % (2 ** 32)
    random.seed(worker_seed)
    np.random.seed(worker_seed)
    torch.manual_seed(worker_seed)

def make_seeded_generator(seed: int, tag: str) -> torch.Generator:
    generator = torch.Generator()
    generator.manual_seed(seed + stable_int_from_text(tag))
    return generator

def cuda_memory_snapshot() -> str:
    if DEVICE.type != "cuda" or not torch.cuda.is_available():
        return "allocated=0.00 GB, reserved=0.00 GB, peak=0.00 GB"
    allocated = torch.cuda.memory_allocated(device=DEVICE) / (1024 ** 3)
    reserved = torch.cuda.memory_reserved(device=DEVICE) / (1024 ** 3)
    peak = torch.cuda.max_memory_allocated(device=DEVICE) / (1024 ** 3)
    return f"allocated={allocated:.2f} GB, reserved={reserved:.2f} GB, peak={peak:.2f} GB"

def run_cuda_cleanup(context: str | None = None) -> None:
    gc.collect()
    if DEVICE.type != "cuda" or not torch.cuda.is_available():
        return
    try:
        torch.cuda.synchronize(device=DEVICE)
    except Exception:
        pass
    try:
        torch.cuda.empty_cache()
    except Exception:
        pass
    try:
        torch.cuda.ipc_collect()
    except Exception:
        pass
    if context is not None:
        print(f"[CUDA Cleanup] {context}: {cuda_memory_snapshot()}")
    try:
        torch.cuda.reset_peak_memory_stats(device=DEVICE)
    except Exception:
        pass

def prune_directory_except(root: Path, keep_file_names: set[str]) -> None:
    if not root.exists():
        return
    keep_paths = {root / name for name in keep_file_names}
    for path in sorted((p for p in root.rglob("*") if p.is_file()), reverse=True):
        if path not in keep_paths:
            path.unlink()
    for path in sorted((p for p in root.rglob("*") if p.is_dir()), reverse=True):
        if path != root:
            try:
                path.rmdir()
            except OSError:
                pass

def prune_optuna_trial_dir(trial_dir: Path) -> None:
    prune_directory_except(trial_dir, {"summary.json"})

def prune_optuna_study_dir(study_root: Path) -> None:
    prune_directory_except(study_root, {"best_params.json", "summary.json", "study.sqlite3"})

def to_device(batch: Any, device: torch.device) -> Any:
    if torch.is_tensor(batch):
        return batch.to(device, non_blocking=True)
    if isinstance(batch, dict):
        return {k: to_device(v, device) for k, v in batch.items()}
    if isinstance(batch, list):
        return [to_device(v, device) for v in batch]
    if isinstance(batch, tuple):
        return tuple(to_device(v, device) for v in batch)
    return batch

def percent_label(percent: float) -> str:
    return f"{int(round(percent * 100)):02d}"

def current_dataset_name() -> str:
    dataset_name = str(DATASET_NAME).strip()
    if dataset_name not in SUPPORTED_DATASET_NAMES:
        raise ValueError(f"DATASET_NAME must be one of {SUPPORTED_DATASET_NAMES}, got {dataset_name!r}")
    return dataset_name

def current_busi_with_classes_split_policy() -> str:
    split_policy = str(BUSI_WITH_CLASSES_SPLIT_POLICY).strip().lower()
    if split_policy not in SUPPORTED_BUSI_WITH_CLASSES_SPLIT_POLICIES:
        raise ValueError(
            f"BUSI_WITH_CLASSES_SPLIT_POLICY must be one of {SUPPORTED_BUSI_WITH_CLASSES_SPLIT_POLICIES}, "
            f"got {split_policy!r}"
        )
    return split_policy

def current_dataset_splits_json_path() -> Path:
    dataset_name = current_dataset_name()
    if dataset_name == "BUSI":
        return DATASET_SPLITS_JSON
    return PROJECT_DIR / f"dataset_splits_{dataset_name.lower()}_{current_busi_with_classes_split_policy()}.json"

def current_dataset_dirs() -> tuple[Path, Path]:
    dataset_name = current_dataset_name()
    if dataset_name == "BUSI":
        return DATA_ROOT / "images", DATA_ROOT / "annotations"
    return DATA_ROOT / "all_images", DATA_ROOT / "all_masks"

def current_pipeline_check_path() -> Path | None:
    if current_dataset_name() != "BUSI_with_classes":
        return None
    return DATA_ROOT / "pipeline_check.json"

def normalization_cache_tag() -> str:
    dataset_name = current_dataset_name()
    if dataset_name == "BUSI":
        return "BUSI"
    return f"{dataset_name}_{current_busi_with_classes_split_policy()}"

def resolve_amp_dtype(key: str) -> torch.dtype:
    key = key.lower().strip()
    if key == "auto":
        if DEVICE.type == "cuda":
            try:
                if torch.cuda.is_bf16_supported():
                    return torch.bfloat16
            except Exception:
                major, _minor = torch.cuda.get_device_capability()
                if major >= 8:
                    return torch.bfloat16
        return torch.float16
    if key in {"float16", "fp16", "half"}:
        return torch.float16
    if key in {"bfloat16", "bf16"}:
        if DEVICE.type == "cuda":
            try:
                if torch.cuda.is_bf16_supported():
                    return torch.bfloat16
            except Exception:
                major, _minor = torch.cuda.get_device_capability()
                if major >= 8:
                    return torch.bfloat16
        print("[AMP] bfloat16 requested but unsupported here. Falling back to float16.")
        return torch.float16
    raise ValueError(f"Unsupported AMP_DTYPE: {key}")

def amp_autocast_enabled(device: torch.device) -> bool:
    return USE_AMP and device.type in {"cuda", "mps"}

def autocast_ctx(enabled: bool, device: torch.device, amp_dtype: torch.dtype):
    if not enabled:
        return nullcontext()
    return torch.autocast(device_type=device.type, dtype=amp_dtype, enabled=True)

def make_grad_scaler(enabled: bool, amp_dtype: torch.dtype, device: torch.device):
    if not enabled or device.type != "cuda" or amp_dtype == torch.bfloat16:
        return None
    try:
        return torch.amp.GradScaler("cuda", enabled=True, init_scale=8192.0)
    except Exception:
        return torch.cuda.amp.GradScaler(enabled=True, init_scale=8192.0)

def format_seconds(seconds: float) -> str:
    seconds = int(seconds)
    h, rem = divmod(seconds, 3600)
    m, s = divmod(rem, 60)
    return f"{h:02d}:{m:02d}:{s:02d}"

def tensor_bytes(t: torch.Tensor) -> int:
    return t.numel() * t.element_size()

def bytes_to_gb(num_bytes: int) -> float:
    return num_bytes / (1024 ** 3)

def set_current_job_params(payload: dict[str, Any] | None = None) -> None:
    CURRENT_JOB_PARAMS.clear()
    if payload:
        CURRENT_JOB_PARAMS.update(dict(payload))

def _job_param(name: str, default: Any) -> Any:
    return CURRENT_JOB_PARAMS.get(name, default)

def _refinement_deltas(*, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
    small = float(_job_param("refine_delta_small", 0.10))
    large = float(_job_param("refine_delta_large", 0.25))
    return torch.tensor([-large, -small, 0.0, small, large], device=device, dtype=dtype)

def threshold_binary_mask(mask: torch.Tensor, threshold: float | None = None) -> torch.Tensor:
    threshold = float(_job_param("threshold", CONTROLLED_MASK_THRESHOLD)) if threshold is None else float(threshold)
    return (mask > threshold).to(dtype=mask.dtype)

def threshold_binary_long(mask: torch.Tensor, threshold: float | None = None) -> torch.Tensor:
    threshold = float(_job_param("threshold", CONTROLLED_MASK_THRESHOLD)) if threshold is None else float(threshold)
    return (mask > threshold).long()

"""=============================================================================
BUSI SPLIT + NORMALIZATION
=============================================================================
"""

IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
IMAGENET_STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)

def validate_image_mask_consistency(images_dir: Path, annotations_dir: Path):
    image_files = {f for f in os.listdir(images_dir) if not f.startswith(".") and f.lower().endswith(".png")}
    mask_files = {f for f in os.listdir(annotations_dir) if not f.startswith(".") and f.lower().endswith(".png")}
    matched = sorted(image_files & mask_files)
    missing_masks = sorted(image_files - mask_files)
    missing_images = sorted(mask_files - image_files)
    return matched, missing_masks, missing_images

def parse_busi_with_classes_label(filename: str) -> str:
    upper_name = str(filename).upper()
    if upper_name.endswith("_B.PNG"):
        return "benign"
    if upper_name.endswith("_M.PNG"):
        return "malignant"
    raise ValueError(
        f"BUSI_with_classes filename must end with '_B.png' or '_M.png', got {filename!r}"
    )

def _candidate_report_dicts(payload: dict[str, Any]) -> list[dict[str, Any]]:
    candidates = [payload]
    for key in ("counts", "summary", "dataset", "report", "metadata"):
        value = payload.get(key)
        if isinstance(value, dict):
            candidates.append(value)
    return candidates

def _extract_report_int(payload: dict[str, Any], keys: tuple[str, ...]) -> int | None:
    for candidate in _candidate_report_dicts(payload):
        for key in keys:
            value = candidate.get(key)
            if isinstance(value, bool):
                continue
            if isinstance(value, (int, np.integer)):
                return int(value)
            if isinstance(value, float) and float(value).is_integer():
                return int(value)
    return None

def _extract_report_filenames(payload: dict[str, Any]) -> set[str] | None:
    for candidate in _candidate_report_dicts(payload):
        filenames = candidate.get("filenames")
        if isinstance(filenames, list) and all(isinstance(item, str) for item in filenames):
            return set(filenames)

        pairs = candidate.get("pairs")
        if isinstance(pairs, list):
            extracted = {item["filename"] for item in pairs if isinstance(item, dict) and isinstance(item.get("filename"), str)}
            if extracted:
                return extracted
    return None

def validate_busi_with_classes_pipeline_report(report_path: Path, sample_records: list[dict[str, str]]) -> None:
    if not report_path.exists():
        return

    payload = load_json(report_path)
    if not isinstance(payload, dict):
        raise RuntimeError(f"Expected dict payload in {report_path}, found {type(payload).__name__}.")

    benign_count = sum(1 for record in sample_records if record.get("class_label") == "benign")
    malignant_count = sum(1 for record in sample_records if record.get("class_label") == "malignant")
    expected_counts = {
        "total_pairs": len(sample_records),
        "benign": benign_count,
        "malignant": malignant_count,
    }
    report_counts = {
        "total_pairs": _extract_report_int(payload, ("total_pairs", "pair_count", "num_pairs", "total")),
        "benign": _extract_report_int(payload, ("benign", "benign_count", "num_benign")),
        "malignant": _extract_report_int(payload, ("malignant", "malignant_count", "num_malignant")),
    }
    for key, expected_value in expected_counts.items():
        report_value = report_counts[key]
        if report_value is not None and report_value != expected_value:
            raise RuntimeError(
                f"pipeline_check mismatch for {key}: discovered={expected_value}, report={report_value} ({report_path})"
            )

    report_filenames = _extract_report_filenames(payload)
    if report_filenames is not None:
        discovered_filenames = {record["filename"] for record in sample_records}
        if report_filenames != discovered_filenames:
            missing_from_report = sorted(discovered_filenames - report_filenames)[:10]
            extra_in_report = sorted(report_filenames - discovered_filenames)[:10]
            raise RuntimeError(
                f"pipeline_check filenames mismatch for {report_path}: "
                f"missing_from_report={missing_from_report}, extra_in_report={extra_in_report}"
            )

    print(f"[Pipeline Check] Validated BUSI_with_classes metadata from {report_path}")

def check_data_leakage(splits: dict[str, list[str]]) -> dict[str, list[str]]:
    leaks: dict[str, list[str]] = {}
    split_names = list(splits.keys())
    for i, lhs in enumerate(split_names):
        for rhs in split_names[i + 1 :]:
            overlap = sorted(set(splits[lhs]) & set(splits[rhs]))
            if overlap:
                leaks[f"{lhs} ∩ {rhs}"] = overlap
    return leaks

def _project_relative_path(path: Path) -> str:
    resolved = Path(path).resolve()
    try:
        return str(resolved.relative_to(PROJECT_DIR.resolve()))
    except ValueError:
        return str(resolved)

def resolve_dataset_root_from_registry(split_registry: dict[str, Any]) -> Path:
    dataset_root = Path(split_registry["dataset_root"])
    if dataset_root.is_absolute():
        return dataset_root
    return (PROJECT_DIR / dataset_root).resolve()

def make_sample_record(
    filename: str,
    images_subdir: str,
    annotations_subdir: str,
    *,
    class_label: str | None = None,
) -> dict[str, str]:
    record = {
        "filename": filename,
        "image_rel_path": str(Path(images_subdir) / filename),
        "mask_rel_path": str(Path(annotations_subdir) / filename),
    }
    if class_label is not None:
        record["class_label"] = class_label
    return record

def build_sample_records(
    filenames: list[str],
    *,
    images_subdir: str,
    annotations_subdir: str,
    dataset_name: str,
) -> list[dict[str, str]]:
    records = []
    for filename in sorted(filenames):
        class_label = parse_busi_with_classes_label(filename) if dataset_name == "BUSI_with_classes" else None
        records.append(
            make_sample_record(
                filename,
                images_subdir,
                annotations_subdir,
                class_label=class_label,
            )
        )
    return records

def split_ratios_for_type(split_type: str) -> tuple[float, float]:
    if split_type == "80_10_10":
        return 0.80, 0.10
    if split_type == "70_10_20":
        return 0.70, 0.10
    raise ValueError(f"Unsupported split_type: {split_type}")

def deterministic_shuffle_records(records: list[dict[str, str]], *, seed: int, tag: str) -> list[dict[str, str]]:
    rng = random.Random(seed + stable_int_from_text(tag))
    shuffled = [dict(record) for record in records]
    rng.shuffle(shuffled)
    return shuffled

def group_records_by_class(sample_records: list[dict[str, str]]) -> dict[str, list[dict[str, str]]]:
    grouped: dict[str, list[dict[str, str]]] = {}
    for record in sample_records:
        class_label = record.get("class_label")
        if class_label is None:
            raise RuntimeError("Expected class_label in sample record for class-aware splitting.")
        grouped.setdefault(class_label, []).append(dict(record))
    return grouped

def allocate_counts_by_ratio(total_size: int, available_counts: dict[str, int]) -> dict[str, int]:
    allocation = {label: 0 for label in available_counts}
    if total_size <= 0 or not available_counts:
        return allocation

    total_available = sum(available_counts.values())
    if total_available <= 0:
        return allocation

    exact = {label: total_size * available_counts[label] / total_available for label in available_counts}
    for label in available_counts:
        allocation[label] = min(available_counts[label], int(math.floor(exact[label])))

    remaining = min(total_size, total_available) - sum(allocation.values())
    order = sorted(
        available_counts.keys(),
        key=lambda label: (exact[label] - math.floor(exact[label]), available_counts[label], label),
        reverse=True,
    )
    while remaining > 0:
        progressed = False
        for label in order:
            if allocation[label] < available_counts[label]:
                allocation[label] += 1
                remaining -= 1
                progressed = True
                if remaining == 0:
                    break
        if not progressed:
            break
    return allocation

def allocate_balanced_counts(total_size: int, available_counts: dict[str, int]) -> dict[str, int]:
    allocation = {label: 0 for label in available_counts}
    if total_size <= 0 or not available_counts:
        return allocation

    labels = sorted(available_counts.keys())
    half = total_size // 2
    for label in labels:
        allocation[label] = min(available_counts[label], half)

    remaining = min(total_size, sum(available_counts.values())) - sum(allocation.values())
    while remaining > 0:
        candidates = [label for label in labels if allocation[label] < available_counts[label]]
        if not candidates:
            break
        best_label = max(
            candidates,
            key=lambda label: (
                available_counts[label] - allocation[label],
                1 if label == "benign" else 0,
                label,
            ),
        )
        allocation[best_label] += 1
        remaining -= 1
    return allocation

def build_unstratified_base_split(
    sample_records: list[dict[str, str]],
    *,
    split_type: str,
    seed: int,
) -> dict[str, list[dict[str, str]]]:
    train_ratio, val_ratio = split_ratios_for_type(split_type)
    records = deterministic_shuffle_records(sample_records, seed=seed, tag=f"base::{split_type}")

    total_samples = len(records)
    train_end = int(total_samples * train_ratio)
    val_end = int(total_samples * (train_ratio + val_ratio))
    return {
        "train": records[:train_end],
        "val": records[train_end:val_end],
        "test": records[val_end:],
    }

def build_stratified_base_split(
    sample_records: list[dict[str, str]],
    *,
    split_type: str,
    seed: int,
) -> dict[str, list[dict[str, str]]]:
    train_ratio, val_ratio = split_ratios_for_type(split_type)
    grouped = group_records_by_class(sample_records)
    splits = {"train": [], "val": [], "test": []}

    for class_label in sorted(grouped.keys()):
        records = deterministic_shuffle_records(
            grouped[class_label],
            seed=seed,
            tag=f"base::{split_type}::{class_label}",
        )
        total_samples = len(records)
        train_end = int(total_samples * train_ratio)
        val_end = int(total_samples * (train_ratio + val_ratio))
        splits["train"].extend(records[:train_end])
        splits["val"].extend(records[train_end:val_end])
        splits["test"].extend(records[val_end:])

    for split_name in splits:
        splits[split_name] = deterministic_shuffle_records(
            splits[split_name],
            seed=seed,
            tag=f"base::{split_type}::{split_name}",
        )
    return splits

def build_balanced_train_base_split(
    sample_records: list[dict[str, str]],
    *,
    split_type: str,
    seed: int,
) -> dict[str, list[dict[str, str]]]:
    train_ratio, val_ratio = split_ratios_for_type(split_type)
    test_ratio = 1.0 - train_ratio - val_ratio
    grouped = group_records_by_class(sample_records)
    if sorted(grouped.keys()) != ["benign", "malignant"]:
        raise RuntimeError(
            f"balanced_train split policy expects benign/malignant classes, found {sorted(grouped.keys())}"
        )

    shuffled = {
        class_label: deterministic_shuffle_records(
            records,
            seed=seed,
            tag=f"base::{split_type}::balanced_train::{class_label}",
        )
        for class_label, records in grouped.items()
    }

    nominal_train_size = int(len(sample_records) * train_ratio)
    per_class_train = min(
        nominal_train_size // 2,
        *(len(records) for records in shuffled.values()),
    )

    train_records: list[dict[str, str]] = []
    remaining_by_class: dict[str, list[dict[str, str]]] = {}
    for class_label in sorted(shuffled.keys()):
        records = shuffled[class_label]
        train_records.extend(records[:per_class_train])
        remaining_by_class[class_label] = records[per_class_train:]

    remainder_val_fraction = val_ratio / max(val_ratio + test_ratio, 1e-8)
    val_records: list[dict[str, str]] = []
    test_records: list[dict[str, str]] = []
    for class_label in sorted(remaining_by_class.keys()):
        records = remaining_by_class[class_label]
        val_count = int(len(records) * remainder_val_fraction)
        val_records.extend(records[:val_count])
        test_records.extend(records[val_count:])

    return {
        "train": deterministic_shuffle_records(
            train_records,
            seed=seed,
            tag=f"base::{split_type}::balanced_train::train",
        ),
        "val": deterministic_shuffle_records(
            val_records,
            seed=seed,
            tag=f"base::{split_type}::balanced_train::val",
        ),
        "test": deterministic_shuffle_records(
            test_records,
            seed=seed,
            tag=f"base::{split_type}::balanced_train::test",
        ),
    }

def build_nested_train_subsets(
    train_records: list[dict[str, str]],
    train_fractions: list[float],
    *,
    split_type: str,
    seed: int,
    split_policy: str | None = None,
) -> dict[str, list[dict[str, str]]]:
    if not train_records:
        return {}

    ordered_records = deterministic_shuffle_records(train_records, seed=seed, tag=f"subset::{split_type}")
    use_class_labels = any("class_label" in record for record in train_records)
    if not use_class_labels:
        subsets: dict[str, list[dict[str, str]]] = {}
        for fraction in sorted({float(value) for value in train_fractions}):
            if fraction <= 0.0 or fraction > 1.0:
                raise ValueError(f"Invalid training fraction: {fraction}")
            subset_key = percent_label(fraction)
            subset_size = len(ordered_records) if fraction >= 1.0 else max(1, int(len(ordered_records) * fraction))
            subsets[subset_key] = [dict(record) for record in ordered_records[:subset_size]]
        return subsets

    grouped = {
        class_label: deterministic_shuffle_records(
            records,
            seed=seed,
            tag=f"subset::{split_type}::{split_policy or 'stratified'}::{class_label}",
        )
        for class_label, records in group_records_by_class(train_records).items()
    }
    available_counts = {class_label: len(records) for class_label, records in grouped.items()}

    subsets: dict[str, list[dict[str, str]]] = {}
    for fraction in sorted({float(value) for value in train_fractions}):
        if fraction <= 0.0 or fraction > 1.0:
            raise ValueError(f"Invalid training fraction: {fraction}")
        subset_key = percent_label(fraction)
        subset_size = len(ordered_records) if fraction >= 1.0 else max(1, int(len(ordered_records) * fraction))
        if split_policy == "balanced_train":
            class_counts = allocate_balanced_counts(subset_size, available_counts)
        else:
            class_counts = allocate_counts_by_ratio(subset_size, available_counts)

        subset_records: list[dict[str, str]] = []
        for class_label in sorted(grouped.keys()):
            subset_records.extend([dict(record) for record in grouped[class_label][: class_counts[class_label]]])
        subsets[subset_key] = deterministic_shuffle_records(
            subset_records,
            seed=seed,
            tag=f"subset::{split_type}::{split_policy or 'stratified'}::{subset_key}",
        )
    return subsets

def train_fraction_from_subset_key(subset_key: str) -> float:
    subset_text = str(subset_key).strip()
    if not subset_text.isdigit():
        raise RuntimeError(f"Invalid train subset key {subset_key!r} in dataset_splits.json.")

    percent = int(subset_text)
    if percent <= 0 or percent > 100:
        raise RuntimeError(f"Train subset key {subset_key!r} must be between 1 and 100.")
    return percent / 100.0

def validate_persisted_split_no_leakage(split_type: str, split_entry: dict[str, Any], *, source: str) -> None:
    base_splits = split_entry["base_splits"]
    base_filenames: dict[str, list[str]] = {}
    for split_name, records in base_splits.items():
        filenames = [record["filename"] for record in records]
        if len(filenames) != len(set(filenames)):
            raise RuntimeError(f"Duplicate filenames detected inside {split_name} for split_type={split_type}.")
        base_filenames[split_name] = filenames

    leaks = check_data_leakage(base_filenames)
    if leaks:
        raise RuntimeError(f"Data leakage detected for split_type={split_type}: {list(leaks.keys())}")

    base_train = set(base_filenames["train"])
    previous_subset: set[str] = set()
    for subset_key in sorted(split_entry["train_subsets"].keys(), key=int):
        subset_filenames = [record["filename"] for record in split_entry["train_subsets"][subset_key]]
        if len(subset_filenames) != len(set(subset_filenames)):
            raise RuntimeError(
                f"Duplicate filenames detected inside train subset {subset_key} for split_type={split_type}."
            )
        subset_set = set(subset_filenames)
        missing = sorted(subset_set - base_train)
        if missing:
            raise RuntimeError(
                f"Train subset {subset_key} contains files outside the base train split for split_type={split_type}."
            )
        if previous_subset and not previous_subset.issubset(subset_set):
            raise RuntimeError(
                f"Train subsets are not nested for split_type={split_type}."
            )
        previous_subset = subset_set

    print(f"[Split Check] No data leakage detected for split_type={split_type} ({source}).")

def repair_persisted_train_subsets(
    split_registry: dict[str, Any],
    requested_train_fractions: list[float],
    *,
    split_json_path: Path,
    seed: int,
) -> bool:
    split_entries = split_registry.get("split_types", {})
    requested_fractions = sorted({float(value) for value in requested_train_fractions})
    combined_fractions = {float(value) for value in split_registry.get("train_fractions", [])}
    combined_fractions.update(requested_fractions)
    dataset_name = str(split_registry.get("dataset_name", "BUSI"))
    split_policy = split_registry.get("split_policy") if dataset_name == "BUSI_with_classes" else None

    for split_type in SUPPORTED_SPLIT_TYPES:
        split_entry = split_entries[split_type]
        train_subsets = split_entry.get("train_subsets", {})
        for subset_key in train_subsets.keys():
            combined_fractions.add(train_fraction_from_subset_key(subset_key))

    combined_fractions_list = sorted(combined_fractions)
    requested_keys = {percent_label(fraction) for fraction in requested_fractions}
    registry_seed = int(split_registry.get("seed", seed))
    repaired = False

    if split_registry.get("train_fractions") != combined_fractions_list:
        split_registry["train_fractions"] = combined_fractions_list
        repaired = True

    for split_type in SUPPORTED_SPLIT_TYPES:
        split_entry = split_entries[split_type]
        train_subsets = split_entry.get("train_subsets", {})
        missing_requested_keys = sorted(requested_keys - set(train_subsets.keys()), key=int)
        if not missing_requested_keys:
            continue

        split_entry["train_subsets"] = build_nested_train_subsets(
            split_entry["base_splits"]["train"],
            combined_fractions_list,
            split_type=split_type,
            seed=registry_seed,
            split_policy=split_policy,
        )
        print(
            f"[Splits] Rebuilt missing train subsets {missing_requested_keys} "
            f"for split_type={split_type} in {split_json_path}"
        )
        repaired = True

    if repaired:
        save_json(split_json_path, split_registry)
        print(f"[Splits] Updated persisted dataset splits at {split_json_path}")
    return repaired

def load_or_create_dataset_splits(
    images_dir: Path,
    annotations_dir: Path,
    split_json_path: Path,
    train_fractions: list[float],
    seed: int,
) -> tuple[dict[str, Any], str]:
    images_dir = Path(images_dir).resolve()
    annotations_dir = Path(annotations_dir).resolve()
    split_json_path = Path(split_json_path).resolve()
    dataset_name = current_dataset_name()
    split_policy = current_busi_with_classes_split_policy() if dataset_name == "BUSI_with_classes" else None
    if split_json_path.exists():
        split_registry = load_json(split_json_path)
        if split_registry.get("version") != DATASET_SPLITS_VERSION:
            raise RuntimeError(
                f"Unsupported dataset_splits.json version in {split_json_path}. "
                f"Expected version={DATASET_SPLITS_VERSION}."
            )
        persisted_dataset_name = str(split_registry.get("dataset_name", "BUSI"))
        if persisted_dataset_name != dataset_name:
            raise RuntimeError(
                f"dataset_splits.json at {split_json_path} targets dataset_name={persisted_dataset_name!r}, "
                f"but current DATASET_NAME={dataset_name!r}."
            )
        persisted_split_policy = split_registry.get("split_policy")
        if dataset_name == "BUSI_with_classes" and persisted_split_policy != split_policy:
            raise RuntimeError(
                f"dataset_splits.json at {split_json_path} targets split_policy={persisted_split_policy!r}, "
                f"but current BUSI_WITH_CLASSES_SPLIT_POLICY={split_policy!r}."
            )
        split_entries = split_registry.get("split_types")
        if not isinstance(split_entries, dict):
            raise RuntimeError(f"Invalid split_types payload in {split_json_path}.")
        for split_type in SUPPORTED_SPLIT_TYPES:
            if split_type not in split_entries:
                raise RuntimeError(
                    f"dataset_splits.json is missing split_type={split_type}. Delete it to regenerate cleanly."
                )
        repaired = repair_persisted_train_subsets(
            split_registry,
            train_fractions,
            split_json_path=split_json_path,
            seed=seed,
        )
        source = "repaired" if repaired else "loaded"
        if dataset_name == "BUSI_with_classes":
            sample_records = build_sample_records(
                validate_image_mask_consistency(images_dir, annotations_dir)[0],
                images_subdir=split_registry["images_subdir"],
                annotations_subdir=split_registry["annotations_subdir"],
                dataset_name=dataset_name,
            )
            pipeline_check_path = current_pipeline_check_path()
            if pipeline_check_path is not None:
                validate_busi_with_classes_pipeline_report(pipeline_check_path, sample_records)
        for split_type in SUPPORTED_SPLIT_TYPES:
            validate_persisted_split_no_leakage(split_type, split_entries[split_type], source=source)
        if repaired:
            print(f"[Splits] Loaded and repaired persisted dataset splits from {split_json_path}")
        else:
            print(f"[Splits] Loaded persisted dataset splits from {split_json_path}")
        return split_registry, source

    matched, missing_masks, missing_images = validate_image_mask_consistency(images_dir, annotations_dir)
    if missing_masks or missing_images:
        raise RuntimeError(
            "BUSI image/mask mismatch detected. "
            f"missing_masks={len(missing_masks)}, missing_images={len(missing_images)}"
        )

    dataset_root = images_dir.parent.resolve()
    images_subdir = images_dir.relative_to(dataset_root).as_posix()
    annotations_subdir = annotations_dir.relative_to(dataset_root).as_posix()
    sample_records = build_sample_records(
        matched,
        images_subdir=images_subdir,
        annotations_subdir=annotations_subdir,
        dataset_name=dataset_name,
    )
    if dataset_name == "BUSI_with_classes":
        pipeline_check_path = current_pipeline_check_path()
        if pipeline_check_path is not None:
            validate_busi_with_classes_pipeline_report(pipeline_check_path, sample_records)
    split_registry = {
        "version": DATASET_SPLITS_VERSION,
        "dataset_name": dataset_name,
        "split_policy": split_policy,
        "dataset_root": _project_relative_path(dataset_root),
        "images_subdir": images_subdir,
        "annotations_subdir": annotations_subdir,
        "seed": seed,
        "train_fractions": sorted({float(value) for value in train_fractions}),
        "split_types": {},
    }

    for split_type in SUPPORTED_SPLIT_TYPES:
        if dataset_name == "BUSI_with_classes":
            if split_policy == "balanced_train":
                base_splits = build_balanced_train_base_split(
                    sample_records,
                    split_type=split_type,
                    seed=seed,
                )
            else:
                base_splits = build_stratified_base_split(
                    sample_records,
                    split_type=split_type,
                    seed=seed,
                )
        else:
            base_splits = build_unstratified_base_split(
                sample_records,
                split_type=split_type,
                seed=seed,
            )
        train_subsets = build_nested_train_subsets(
            base_splits["train"],
            train_fractions,
            split_type=split_type,
            seed=seed,
            split_policy=split_policy,
        )
        split_entry = {
            "split_type": split_type,
            "base_splits": base_splits,
            "train_subsets": train_subsets,
        }
        validate_persisted_split_no_leakage(split_type, split_entry, source="created")
        split_registry["split_types"][split_type] = split_entry

    save_json(split_json_path, split_registry)
    print(f"[Splits] Created persisted dataset splits at {split_json_path}")
    return split_registry, "created"

def select_persisted_split(
    split_registry: dict[str, Any],
    split_type: str,
    train_fraction: float,
) -> dict[str, Any]:
    if split_type not in SUPPORTED_SPLIT_TYPES:
        raise ValueError(f"Unsupported split_type: {split_type}")

    split_entries = split_registry.get("split_types", {})
    if split_type not in split_entries:
        raise KeyError(
            f"Requested split_type={split_type} is not available in dataset_splits.json. "
            "Delete the JSON file to regenerate it with the new configuration."
        )

    subset_key = percent_label(train_fraction)
    split_entry = split_entries[split_type]
    train_subsets = split_entry.get("train_subsets", {})
    if subset_key not in train_subsets:
        raise KeyError(
            f"Requested train fraction={train_fraction} (key={subset_key}) is not available in dataset_splits.json."
        )

    return {
        "dataset_root": resolve_dataset_root_from_registry(split_registry),
        "split_type": split_type,
        "train_fraction": float(train_fraction),
        "train_subset_key": subset_key,
        "base_train_records": split_entry["base_splits"]["train"],
        "train_records": train_subsets[subset_key],
        "val_records": split_entry["base_splits"]["val"],
        "test_records": split_entry["base_splits"]["test"],
    }

def compute_busi_statistics(
    dataset_root: Path,
    sample_records: list[dict[str, str]],
    cache_path: Path,
) -> tuple[float, float, str]:
    filenames = [record["filename"] for record in sample_records]
    if cache_path.exists():
        stats = load_json(cache_path)
        if stats.get("filenames") == filenames:
            print(f"[Normalization] Loaded cached normalization stats from {cache_path}")
            return float(stats["global_mean"]), float(stats["global_std"]), "loaded_from_cache"

    total_sum = np.float64(0.0)
    total_sq_sum = np.float64(0.0)
    total_pixels = 0

    for record in tqdm(sample_records, desc="Computing BUSI train mean/std", leave=False):
        image_path = dataset_root / record["image_rel_path"]
        img = np.array(PILImage.open(image_path)).astype(np.float64)
        total_sum += img.sum()
        total_sq_sum += (img ** 2).sum()
        total_pixels += img.size

    global_mean = float(total_sum / total_pixels)
    global_std = float(np.sqrt(total_sq_sum / total_pixels - global_mean ** 2))
    if global_std < 1e-6:
        global_std = 1.0

    save_json(
        cache_path,
        {
            "global_mean": global_mean,
            "global_std": global_std,
            "total_pixels": int(total_pixels),
            "num_images": len(sample_records),
            "filenames": filenames,
        },
    )
    print(f"[Normalization] Computed and saved normalization stats to {cache_path}")
    return global_mean, global_std, "computed_fresh"

def compute_class_distribution(sample_records: list[dict[str, str]]) -> dict[str, int] | None:
    if not sample_records or not any("class_label" in record for record in sample_records):
        return None
    return {
        "benign": sum(1 for record in sample_records if record.get("class_label") == "benign"),
        "malignant": sum(1 for record in sample_records if record.get("class_label") == "malignant"),
    }

def format_class_distribution(class_distribution: dict[str, int] | None) -> str:
    if class_distribution is None:
        return "unavailable"
    benign = int(class_distribution.get("benign", 0))
    malignant = int(class_distribution.get("malignant", 0))
    total = benign + malignant
    return f"benign={benign}, malignant={malignant}, total={total}"

def print_loaded_class_distribution(
    *,
    split_type: str,
    train_subset_key: str,
    base_train_records: list[dict[str, str]],
    train_records: list[dict[str, str]],
    val_records: list[dict[str, str]],
    test_records: list[dict[str, str]],
) -> None:
    if not any("class_label" in record for record in train_records):
        return
    section(f"Loaded Class Distribution | {split_type} | {train_subset_key}%")
    print(f"Base train classes    : {format_class_distribution(compute_class_distribution(base_train_records))}")
    print(f"Train subset classes  : {format_class_distribution(compute_class_distribution(train_records))}")
    print(f"Validation classes    : {format_class_distribution(compute_class_distribution(val_records))}")
    print(f"Test classes          : {format_class_distribution(compute_class_distribution(test_records))}")

def print_split_summary(payload: dict[str, Any]) -> None:
    section(f"Split Summary | {payload['split_type']} | {payload['train_subset_key']}%")
    print(f"Dataset name           : {payload['dataset_name']}")
    if payload.get("dataset_split_policy") is not None:
        print(f"Dataset split policy   : {payload['dataset_split_policy']}")
    print(f"Dataset splits JSON    : {payload['dataset_splits_path']}")
    print(f"Split source           : {payload['split_source']}")
    print(f"Split type used        : {payload['split_type']}")
    print(f"Train fraction         : {payload['train_subset_key']}% of frozen base train")
    print(f"Base train samples     : {payload['base_train_count']}")
    print(f"Train subset samples   : {payload['train_count']}")
    print(f"Validation samples     : {payload['val_count']}")
    print(f"Test samples           : {payload['test_count']}")
    if payload.get("base_train_class_distribution") is not None:
        print(f"Base train classes     : {format_class_distribution(payload['base_train_class_distribution'])}")
        print(f"Train subset classes   : {format_class_distribution(payload['train_class_distribution'])}")
        print(f"Validation classes     : {format_class_distribution(payload['val_class_distribution'])}")
        print(f"Test classes           : {format_class_distribution(payload['test_class_distribution'])}")
    print(f"Validation/Test frozen : {payload['val_test_frozen']}")
    print(f"Leakage check          : {payload['leakage_check']}")

def print_normalization_summary(payload: dict[str, Any]) -> None:
    mode = "ImageNet mean/std" if USE_IMAGENET_NORM else "Dataset train mean/std"
    print(f"Dataset name           : {payload['dataset_name']}")
    print(f"Normalization mode     : {mode}")
    print(f"Stats cache path       : {payload['normalization_cache_path']}")
    print(f"Stats source           : {payload['normalization_source']}")
    print(f"Split type used        : {payload['split_type']}")
    print(f"Stats computed from    : {payload['train_count']} train samples ({payload['train_subset_key']}%)")
# =============================================================================
# IMAGE PREPARATION + DATASETS
# =============================================================================

def _to_three_channels(image: np.ndarray) -> np.ndarray:
    if image.ndim == 2:
        image = image[..., None]
    if image.shape[2] == 1:
        image = np.repeat(image, 3, axis=2)
    elif image.shape[2] > 3:
        image = image[..., :3]
    return image

def _prepare_image(raw: np.ndarray, global_mean: float, global_std: float) -> np.ndarray:
    img = raw.astype(np.float32)
    img = _to_three_channels(img)
    if IMG_SIZE > 0 and (img.shape[0] != IMG_SIZE or img.shape[1] != IMG_SIZE):
        img = np.array(
            PILImage.fromarray(img.astype(np.uint8)).resize((IMG_SIZE, IMG_SIZE), PILImage.BILINEAR)
        ).astype(np.float32)
    if USE_IMAGENET_NORM:
        if img.max() > 1.0:
            img = img / 255.0
        img = (img - IMAGENET_MEAN) / IMAGENET_STD
    else:
        img = (img - global_mean) / global_std
    return np.transpose(img, (2, 0, 1)).copy()

def _prepare_mask(raw: np.ndarray) -> np.ndarray:
    mask = raw.astype(np.uint8)
    if mask.ndim == 3:
        mask = mask[..., 0]
    if IMG_SIZE > 0 and (mask.shape[0] != IMG_SIZE or mask.shape[1] != IMG_SIZE):
        pil_mask = PILImage.fromarray(mask)
        if pil_mask.mode != "L":
            pil_mask = pil_mask.convert("L")
        mask = np.array(pil_mask.resize((IMG_SIZE, IMG_SIZE), PILImage.NEAREST))
    return ((mask > 0).astype(np.float32))[None, ...].copy()

def print_imagenet_normalization_status() -> bool:
    uses_imagenet_norm = bool(USE_IMAGENET_NORM)
    if uses_imagenet_norm:
        print("✅🖼️ ImageNet normalization is ACTIVE in `_prepare_image`.")
    else:
        print("⚠️🧪 ImageNet normalization is NOT active in `_prepare_image`.")
        print("⚠️📊 Using dataset global mean/std normalization instead.")
    if SMP_ENCODER_WEIGHTS == "imagenet" and not uses_imagenet_norm:
        print("⚠️🚨 Encoder weights are set to ImageNet, but ImageNet normalization is disabled.")
    return uses_imagenet_norm

def _apply_minimal_train_aug(image: torch.Tensor, mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    if torch.rand(1).item() < 0.5:
        image = torch.flip(image, dims=(2,))
        mask = torch.flip(mask, dims=(2,))
    if torch.rand(1).item() < 0.5:
        image = torch.flip(image, dims=(1,))
        mask = torch.flip(mask, dims=(1,))
    if torch.rand(1).item() < 0.5:
        k = 1 if torch.rand(1).item() < 0.5 else 3
        image = torch.rot90(image, k=k, dims=(1, 2))
        mask = torch.rot90(mask, k=k, dims=(1, 2))
    return image.contiguous(), mask.contiguous()

class BUSIDataset(Dataset):
    def __init__(
        self,
        sample_records: list[dict[str, str]],
        dataset_root: Path,
        global_mean: float,
        global_std: float,
        *,
        preload: bool,
        augment: bool,
        split_name: str,
    ) -> None:
        super().__init__()
        self.sample_records = [dict(record) for record in sample_records]
        self.dataset_root = Path(dataset_root)
        self.global_mean = float(global_mean)
        self.global_std = float(global_std)
        self.preload = preload
        self.augment = augment
        self.split_name = split_name
        self._images: list[torch.Tensor] = []
        self._masks: list[torch.Tensor] = []
        self._raw_cache_bytes = 0

        if not self.preload:
            raise ValueError("PRELOAD_TO_RAM is mandatory in this RunPod runner.")
        self._preload_to_ram()

    def _preload_to_ram(self) -> None:
        desc = f"Preloading {self.split_name} ({len(self.sample_records)} samples) to RAM"
        for record in tqdm(self.sample_records, desc=desc, leave=False):
            raw_img = np.array(PILImage.open(self.dataset_root / record["image_rel_path"]))
            raw_mask = np.array(PILImage.open(self.dataset_root / record["mask_rel_path"]))
            if raw_img.shape[:2] != raw_mask.shape[:2]:
                raise RuntimeError(
                    f"Image/mask spatial size mismatch for {record['filename']}: "
                    f"image={raw_img.shape[:2]}, mask={raw_mask.shape[:2]}"
                )
            image = torch.from_numpy(_prepare_image(raw_img, self.global_mean, self.global_std))
            mask = torch.from_numpy(_prepare_mask(raw_mask))
            self._raw_cache_bytes += tensor_bytes(image) + tensor_bytes(mask)
            self._images.append(image)
            self._masks.append(mask)

    def __len__(self) -> int:
        return len(self.sample_records)

    def __getitem__(self, index: int) -> dict[str, Any]:
        image = self._images[index].clone()
        mask = self._masks[index].clone()
        if self.augment:
            image, mask = _apply_minimal_train_aug(image, mask)
        return {
            "image": image,
            "mask": mask,
            "sample_id": Path(self.sample_records[index]["filename"]).stem,
            "dataset": current_dataset_name(),
        }

    @property
    def cache_bytes(self) -> int:
        return self._raw_cache_bytes

class CUDAPrefetcher:
    def __init__(self, loader: DataLoader, device: torch.device) -> None:
        self.loader = loader
        self.device = device
        self._use_cuda = device.type == "cuda"
        self._iter = None
        self._stream = None
        self._next_batch = None

    def __len__(self) -> int:
        return len(self.loader)

    def __iter__(self):
        self._iter = iter(self.loader)
        self._stream = torch.cuda.Stream(device=self.device) if self._use_cuda else None
        self._next_batch = None
        self._preload()
        return self

    def close(self) -> None:
        self._next_batch = None
        self._iter = None
        self._stream = None

    def _preload(self) -> None:
        if self._iter is None:
            self._next_batch = None
            return
        try:
            self._next_batch = next(self._iter)
        except StopIteration:
            self._next_batch = None
            return
        if self._use_cuda:
            assert self._stream is not None
            with torch.cuda.stream(self._stream):
                self._next_batch = to_device(self._next_batch, self.device)
        else:
            self._next_batch = to_device(self._next_batch, self.device)

    def __next__(self):
        if self._next_batch is None:
            self.close()
            raise StopIteration
        if self._use_cuda:
            assert self._stream is not None
            torch.cuda.current_stream(self.device).wait_stream(self._stream)
        batch = self._next_batch
        self._preload()
        if self._next_batch is None:
            self._iter = None
            self._stream = None
        return batch

class DataBundle:
    def __init__(
        self,
        *,
        percent: float,
        split_payload: dict[str, Any],
        train_ds: BUSIDataset,
        val_ds: BUSIDataset,
        test_ds: BUSIDataset,
        train_loader: DataLoader,
        val_loader: DataLoader,
        test_loader: DataLoader,
    ) -> None:
        self.percent = percent
        self.split_payload = split_payload
        self.train_ds = train_ds
        self.val_ds = val_ds
        self.test_ds = test_ds
        self.train_loader = train_loader
        self.val_loader = val_loader
        self.test_loader = test_loader

    @property
    def global_mean(self) -> float:
        return float(self.split_payload["global_mean"])

    @property
    def global_std(self) -> float:
        return float(self.split_payload["global_std"])

    @property
    def total_cache_bytes(self) -> int:
        return self.train_ds.cache_bytes + self.val_ds.cache_bytes + self.test_ds.cache_bytes

def make_loader(dataset: Dataset, shuffle: bool, *, loader_tag: str) -> DataLoader:
    num_workers = NUM_WORKERS
    persistent_workers = USE_PERSISTENT_WORKERS and num_workers > 0
    pin_memory = USE_PIN_MEMORY and DEVICE.type == "cuda"
    generator = make_seeded_generator(SEED, loader_tag)
    return DataLoader(
        dataset,
        batch_size=BATCH_SIZE,
        shuffle=shuffle,
        num_workers=num_workers,
        pin_memory=pin_memory,
        drop_last=False,
        persistent_workers=persistent_workers,
        worker_init_fn=seed_worker,
        generator=generator,
    )

def build_data_bundle(percent: float, split_registry: dict[str, Any], split_source: str) -> DataBundle:
    pct_label = percent_label(percent)
    pct_root = ensure_dir(RUNS_ROOT / MODEL_NAME / f"pct_{pct_label}")
    stats_cache_path = pct_root / f"norm_stats_{normalization_cache_tag()}_{SPLIT_TYPE}_{pct_label}pct.json"

    selected_split = select_persisted_split(split_registry, SPLIT_TYPE, percent)
    base_train_class_distribution = compute_class_distribution(selected_split["base_train_records"])
    train_class_distribution = compute_class_distribution(selected_split["train_records"])
    val_class_distribution = compute_class_distribution(selected_split["val_records"])
    test_class_distribution = compute_class_distribution(selected_split["test_records"])
    print_loaded_class_distribution(
        split_type=selected_split["split_type"],
        train_subset_key=selected_split["train_subset_key"],
        base_train_records=selected_split["base_train_records"],
        train_records=selected_split["train_records"],
        val_records=selected_split["val_records"],
        test_records=selected_split["test_records"],
    )
    dataset_root = Path(selected_split["dataset_root"]).resolve()
    global_mean, global_std, normalization_source = compute_busi_statistics(
        dataset_root=dataset_root,
        sample_records=selected_split["train_records"],
        cache_path=stats_cache_path,
    )

    split_payload = {
        "dataset_name": current_dataset_name(),
        "dataset_split_policy": split_registry.get("split_policy"),
        "dataset_splits_path": str(current_dataset_splits_json_path().resolve()),
        "dataset_root": str(dataset_root),
        "split_source": split_source,
        "split_type": SPLIT_TYPE,
        "dataset_percent": percent,
        "train_subset_key": selected_split["train_subset_key"],
        "base_train_count": len(selected_split["base_train_records"]),
        "train_count": len(selected_split["train_records"]),
        "val_count": len(selected_split["val_records"]),
        "test_count": len(selected_split["test_records"]),
        "base_train_class_distribution": base_train_class_distribution,
        "train_class_distribution": train_class_distribution,
        "val_class_distribution": val_class_distribution,
        "test_class_distribution": test_class_distribution,
        "val_test_frozen": True,
        "leakage_check": "passed",
        "global_mean": global_mean,
        "global_std": global_std,
        "normalization_cache_path": str(stats_cache_path.resolve()),
        "normalization_source": normalization_source,
    }

    print_split_summary(split_payload)
    print_normalization_summary(split_payload)

    train_ds = BUSIDataset(
        selected_split["train_records"],
        dataset_root,
        global_mean,
        global_std,
        preload=PRELOAD_TO_RAM,
        augment=True,
        split_name=f"train {SPLIT_TYPE} {pct_label}%",
    )
    val_ds = BUSIDataset(
        selected_split["val_records"],
        dataset_root,
        global_mean,
        global_std,
        preload=PRELOAD_TO_RAM,
        augment=False,
        split_name=f"val {SPLIT_TYPE}",
    )
    test_ds = BUSIDataset(
        selected_split["test_records"],
        dataset_root,
        global_mean,
        global_std,
        preload=PRELOAD_TO_RAM,
        augment=False,
        split_name=f"test {SPLIT_TYPE}",
    )

    bundle = DataBundle(
        percent=percent,
        split_payload=split_payload,
        train_ds=train_ds,
        val_ds=val_ds,
        test_ds=test_ds,
        train_loader=make_loader(train_ds, shuffle=True, loader_tag=f"{SPLIT_TYPE}:{pct_label}:train"),
        val_loader=make_loader(val_ds, shuffle=False, loader_tag=f"{SPLIT_TYPE}:{pct_label}:val"),
        test_loader=make_loader(test_ds, shuffle=False, loader_tag=f"{SPLIT_TYPE}:{pct_label}:test"),
    )
    print_preload_summary(bundle)
    return bundle

def print_preload_summary(bundle: DataBundle) -> None:
    section(
        f"RAM Preload Summary | {bundle.split_payload['split_type']} | {int(bundle.percent * 100)}%"
    )
    print(f"Train samples         : {len(bundle.train_ds)}")
    print(f"Val samples           : {len(bundle.val_ds)}")
    print(f"Test samples          : {len(bundle.test_ds)}")
    print(f"Train batches         : {len(bundle.train_loader)}")
    print(f"Val batches           : {len(bundle.val_loader)}")
    print(f"Test batches          : {len(bundle.test_loader)}")
    print(f"Global mean           : {bundle.global_mean:.6f}")
    print(f"Global std            : {bundle.global_std:.6f}")
    first = bundle.train_ds[0]
    print(f"Sample image shape    : {tuple(first['image'].shape)}")
    print(f"Sample mask shape     : {tuple(first['mask'].shape)}")
    print(f"Sample image dtype    : {first['image'].dtype}")
    print(f"Sample mask dtype     : {first['mask'].dtype}")
    print(f"Estimated RAM preload : {bytes_to_gb(bundle.total_cache_bytes):.3f} GB")

"""=============================================================================
MODEL DEFINITIONS
=============================================================================
"""

def strategy_name(strategy: int, model_config: RuntimeModelConfig | None = None) -> str:
    model_config = (model_config or current_model_config()).validate()
    if model_config.backbone_family == "custom_vgg":
        if strategy == 1:
            return "Strategy 1: Custom VGG + RL"
        if strategy == 2:
            return "Strategy 2: Custom VGG + Segmentation Head (Supervised)"
        if strategy == 3:
            return "Strategy 3: Custom VGG + Segmentation Head + RL"
        if strategy == 4:
            return "Strategy 4: Frozen Custom VGG + Segmentation Head + RL"
        if strategy == 5:
            return "Strategy 5: Frozen Custom VGG + RL"
        raise ValueError(f"Unsupported strategy: {strategy}")

    if strategy == 1:
        return f"Strategy 1: SMP Encoder ({model_config.smp_encoder_name}) + RL"
    if strategy == 2:
        return f"Strategy 2: SMP {model_config.smp_decoder_type} ({model_config.smp_encoder_name}) supervised"
    if strategy == 3:
        return f"Strategy 3: SMP {model_config.smp_decoder_type} ({model_config.smp_encoder_name}) + RL"
    if strategy == 4:
        return f"Strategy 4: Frozen SMP {model_config.smp_decoder_type} ({model_config.smp_encoder_name}) + RL"
    if strategy == 5:
        return f"Strategy 5: Frozen SMP Encoder ({model_config.smp_encoder_name}) + RL"
    raise ValueError(f"Unsupported strategy: {strategy}")

def _conv3x3(in_ch: int, out_ch: int, dilation: int = 1) -> nn.Conv2d:
    return nn.Conv2d(
        in_ch,
        out_ch,
        kernel_size=3,
        stride=1,
        padding=dilation,
        dilation=dilation,
        bias=True,
    )

class _ConvBlock(nn.Module):
    def __init__(
        self,
        in_ch: int,
        out_ch: int,
        dilation: int = 1,
        *,
        num_groups: int = 0,
        dropout: float = 0.0,
    ) -> None:
        super().__init__()
        self.conv = _conv3x3(in_ch, out_ch, dilation=dilation)
        self.norm = _group_norm(out_ch, num_groups=num_groups) if num_groups > 0 else nn.Identity()
        self.act = nn.ReLU(inplace=True)
        self.drop = nn.Dropout2d(p=dropout) if dropout > 0 else nn.Identity()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.drop(self.act(self.norm(self.conv(x))))

def _group_norm(num_channels: int, *, num_groups: int = GN_NUM_GROUPS) -> nn.GroupNorm:
    groups = min(num_groups, num_channels)
    while groups > 1 and num_channels % groups != 0:
        groups -= 1
    return nn.GroupNorm(groups, num_channels)

class SelfAttentionModule(nn.Module):
    def __init__(self, channels: int) -> None:
        super().__init__()
        mid = max(channels // 8, 1)
        self.query = nn.Conv2d(channels, mid, 1)
        self.key = nn.Conv2d(channels, mid, 1)
        self.value = nn.Conv2d(channels, channels, 1)
        self.gamma = nn.Parameter(torch.zeros(1))

    def forward(self, f: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        b, c, h, w = f.shape
        pooled = f
        if h * w > ATTENTION_MAX_TOKENS:
            stride_h = max(1, math.ceil(h / ATTENTION_MIN_POOL_SIZE))
            stride_w = max(1, math.ceil(w / ATTENTION_MIN_POOL_SIZE))
            pooled = F.avg_pool2d(f, kernel_size=(stride_h, stride_w), stride=(stride_h, stride_w))

        ph, pw = pooled.shape[-2:]
        q = self.query(pooled).view(b, -1, ph * pw).permute(0, 2, 1)
        k = self.key(pooled).view(b, -1, ph * pw)
        v = self.value(pooled).view(b, -1, ph * pw).permute(0, 2, 1)
        attn = torch.softmax(q @ k / (q.shape[-1] ** 0.5), dim=-1)
        out = (attn @ v).permute(0, 2, 1).view(b, c, ph, pw)
        if ph != h or pw != w:
            out = F.interpolate(out, size=(h, w), mode="bilinear", align_corners=False)
        return f + self.gamma * out, attn

class DilatedPolicyHead(nn.Module):
    def __init__(self, in_channels: int) -> None:
        super().__init__()
        self.body = nn.Sequential(
            _ConvBlock(in_channels, 512, dilation=1),
            _ConvBlock(512, 256, dilation=2),
            _ConvBlock(256, 128, dilation=3),
            _ConvBlock(128, 64, dilation=4),
        )
        self.classifier = nn.Conv2d(64, NUM_ACTIONS, kernel_size=1)
        nn.init.zeros_(self.classifier.weight)
        bias = torch.full((NUM_ACTIONS,), -2.0, dtype=torch.float32)
        keep_index = NUM_ACTIONS // 2 if NUM_ACTIONS >= 3 else NUM_ACTIONS - 1
        bias[keep_index] = 2.0
        with torch.no_grad():
            self.classifier.bias.copy_(bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.classifier(self.body(x))

class DilatedValueHead(nn.Module):
    def __init__(self, in_channels: int) -> None:
        super().__init__()
        self.net = nn.Sequential(
            _ConvBlock(in_channels, 512, dilation=1),
            _ConvBlock(512, 256, dilation=2),
            _ConvBlock(256, 128, dilation=3),
            _ConvBlock(128, 64, dilation=4),
            nn.Conv2d(64, 1, kernel_size=1),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)

def replace_bn_with_gn(model: nn.Module, num_groups: int = 8) -> nn.Module:
    for name, module in model.named_children():
        if isinstance(module, (nn.BatchNorm2d, nn.BatchNorm1d)):
            num_channels = module.num_features
            groups = min(num_groups, num_channels)
            while groups > 1 and num_channels % groups != 0:
                groups -= 1
            setattr(model, name, nn.GroupNorm(groups, num_channels, eps=module.eps, affine=module.affine))
        else:
            replace_bn_with_gn(module, num_groups=num_groups)
    return model

class HalfVGG16DilatedExtractor(nn.Module):
    def __init__(self, *, dilation: int = 1, num_scales: int = 3) -> None:
        super().__init__()
        self.num_scales = num_scales
        deep_dropout = 0.1

        self.conv1_1 = _ConvBlock(3, 32, dilation=dilation, num_groups=GN_NUM_GROUPS)
        self.conv1_2 = _ConvBlock(32, 32, dilation=dilation, num_groups=GN_NUM_GROUPS)

        self.conv2_1 = _ConvBlock(32, 64, dilation=dilation, num_groups=GN_NUM_GROUPS)
        self.conv2_2 = _ConvBlock(64, 64, dilation=dilation, num_groups=GN_NUM_GROUPS)

        self.conv3_1 = _ConvBlock(64, 128, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)
        self.conv3_2 = _ConvBlock(128, 128, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)
        self.conv3_3 = _ConvBlock(128, 128, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)

        self.conv4_1 = _ConvBlock(128, 256, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)
        self.conv4_2 = _ConvBlock(256, 256, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)
        self.conv4_3 = _ConvBlock(256, 256, dilation=dilation, num_groups=GN_NUM_GROUPS, dropout=deep_dropout)

        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)

    @property
    def out_channels(self) -> int:
        return (32 + 64 + 128) if self.num_scales == 3 else (32 + 64 + 128 + 256)

    @property
    def pyramid_channels(self) -> list[int]:
        return [32, 64, 128] if self.num_scales == 3 else [32, 64, 128, 256]

    def forward_pyramid(self, x: torch.Tensor) -> list[torch.Tensor]:
        x = self.conv1_1(x)
        src1 = self.conv1_2(x)
        x = self.pool(src1)

        x = self.conv2_1(x)
        src2 = self.conv2_2(x)
        x = self.pool(src2)

        x = self.conv3_1(x)
        x = self.conv3_2(x)
        src3 = self.conv3_3(x)

        if self.num_scales == 3:
            return [src1, src2, src3]

        x = self.pool(src3)
        x = self.conv4_1(x)
        x = self.conv4_2(x)
        src4 = self.conv4_3(x)
        return [src1, src2, src3, src4]

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        h, w = x.shape[-2:]
        pyramid = self.forward_pyramid(x)
        upsampled = []
        for feat in pyramid:
            if feat.shape[-2] != h or feat.shape[-1] != w:
                feat = F.interpolate(feat, size=(h, w), mode="bilinear", align_corners=False)
            upsampled.append(feat)
        return torch.cat(upsampled, dim=1)

class CustomVGGEncoderWrapper(nn.Module):
    def __init__(self, *, num_scales: int, dilation: int) -> None:
        super().__init__()
        self.encoder = HalfVGG16DilatedExtractor(dilation=dilation, num_scales=num_scales)
        self.projection = None

    @property
    def out_channels(self) -> int:
        return self.encoder.out_channels

    @property
    def pyramid_channels(self) -> list[int]:
        return self.encoder.pyramid_channels

    def forward_pyramid(self, x: torch.Tensor) -> list[torch.Tensor]:
        return self.encoder.forward_pyramid(x)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.encoder(x)

class SMPEncoderWrapper(nn.Module):
    def __init__(
        self,
        *,
        encoder_name: str,
        encoder_weights: str | None,
        depth: int,
        in_channels: int,
        proj_dim: int,
    ) -> None:
        super().__init__()
        self.encoder = smp.encoders.get_encoder(
            encoder_name,
            in_channels=in_channels,
            depth=depth,
            weights=encoder_weights,
        )
        if REPLACE_BN_WITH_GN:
            replace_bn_with_gn(self.encoder, num_groups=GN_NUM_GROUPS)

        raw_channels = sum(self.encoder.out_channels)
        if proj_dim > 0:
            groups = min(GN_NUM_GROUPS, proj_dim)
            while groups > 1 and proj_dim % groups != 0:
                groups -= 1
            self.projection = nn.Sequential(
                nn.Conv2d(raw_channels, proj_dim, kernel_size=1, bias=False),
                nn.GroupNorm(groups, proj_dim) if REPLACE_BN_WITH_GN else nn.BatchNorm2d(proj_dim),
                nn.ReLU(inplace=True),
            )
            self._out_channels = proj_dim
        else:
            self.projection = None
            self._out_channels = raw_channels

    @property
    def out_channels(self) -> int:
        return self._out_channels

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        h, w = x.shape[-2:]
        features = self.encoder(x)
        upsampled = []
        for feat in features:
            if feat.shape[-2] != h or feat.shape[-1] != w:
                feat = F.interpolate(feat, size=(h, w), mode="bilinear", align_corners=False)
            upsampled.append(feat)
        out = torch.cat(upsampled, dim=1)
        if self.projection is not None:
            out = self.projection(out)
        return out

class VGGDecoderBlock(nn.Module):
    def __init__(self, *, in_channels: int, skip_channels: int, out_channels: int) -> None:
        super().__init__()
        self.block = nn.Sequential(
            _ConvBlock(in_channels + skip_channels, out_channels, num_groups=GN_NUM_GROUPS),
            _ConvBlock(out_channels, out_channels, num_groups=GN_NUM_GROUPS),
        )

    def forward(self, x: torch.Tensor, skip: torch.Tensor) -> torch.Tensor:
        x = F.interpolate(x, size=skip.shape[-2:], mode="bilinear", align_corners=False)
        return self.block(torch.cat([x, skip], dim=1))

class VGGSegmentationHead(nn.Module):
    def __init__(self, *, pyramid_channels: list[int], dropout_p: float) -> None:
        super().__init__()
        if len(pyramid_channels) not in {3, 4}:
            raise ValueError(f"Expected 3 or 4 VGG pyramid channels, got {pyramid_channels}")

        self.dropout = nn.Dropout2d(p=dropout_p)
        self.num_scales = len(pyramid_channels)

        deepest = pyramid_channels[-1]
        self.bridge = _ConvBlock(deepest, deepest, num_groups=GN_NUM_GROUPS)
        if self.num_scales == 4:
            self.up3 = VGGDecoderBlock(in_channels=deepest, skip_channels=pyramid_channels[2], out_channels=128)
            self.up2 = VGGDecoderBlock(in_channels=128, skip_channels=pyramid_channels[1], out_channels=64)
            self.up1 = VGGDecoderBlock(in_channels=64, skip_channels=pyramid_channels[0], out_channels=32)
        else:
            self.up2 = VGGDecoderBlock(in_channels=deepest, skip_channels=pyramid_channels[1], out_channels=64)
            self.up1 = VGGDecoderBlock(in_channels=64, skip_channels=pyramid_channels[0], out_channels=32)
        self.out_conv = nn.Conv2d(32, 1, kernel_size=1)

    def forward(self, pyramid: list[torch.Tensor] | tuple[torch.Tensor, ...]) -> torch.Tensor:
        features = list(pyramid)
        x = self.bridge(self.dropout(features[-1]))
        if self.num_scales == 4:
            x = self.up3(x, features[2])
            x = self.up2(x, features[1])
            x = self.up1(x, features[0])
        else:
            x = self.up2(x, features[1])
            x = self.up1(x, features[0])
        return self.out_conv(x)

class PixelDRLMG_SMP(nn.Module):
    def __init__(
        self,
        *,
        encoder_name: str,
        encoder_weights: str | None,
        encoder_depth: int,
        proj_dim: int,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.extractor = SMPEncoderWrapper(
            encoder_name=encoder_name,
            encoder_weights=encoder_weights,
            depth=encoder_depth,
            in_channels=3,
            proj_dim=proj_dim,
        )
        ch = self.extractor.out_channels
        self.sam = SelfAttentionModule(channels=ch)
        self.policy_head = DilatedPolicyHead(in_channels=ch)
        self.value_head = DilatedValueHead(in_channels=ch)
        self.head_dropout = nn.Dropout2d(p=dropout_p)
        self.omega_conv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
        nn.init.constant_(self.omega_conv.weight, 1.0 / 9.0)

    def forward_state(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        features = self.extractor(x)
        return self.sam(features)

    def forward_from_state(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state)

    def value_from_state(self, state: torch.Tensor) -> torch.Tensor:
        state = self.head_dropout(state)
        return self.value_head(state)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        state, attention = self.forward_state(x)
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state), attention

    def forward_policy_only(self, x: torch.Tensor) -> torch.Tensor:
        state, _ = self.forward_state(x)
        state = self.head_dropout(state)
        return self.policy_head(state)

    def neighborhood_value(self, value_next: torch.Tensor) -> torch.Tensor:
        return self.omega_conv(value_next)

class PixelDRLMG_VGG(nn.Module):
    def __init__(
        self,
        *,
        num_scales: int,
        dilation: int,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.extractor = CustomVGGEncoderWrapper(num_scales=num_scales, dilation=dilation)
        ch = self.extractor.out_channels
        self.sam = SelfAttentionModule(channels=ch)
        self.policy_head = DilatedPolicyHead(in_channels=ch)
        self.value_head = DilatedValueHead(in_channels=ch)
        self.head_dropout = nn.Dropout2d(p=dropout_p)
        self.omega_conv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
        nn.init.constant_(self.omega_conv.weight, 1.0 / 9.0)

    def forward_state(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        return self.sam(self.extractor(x))

    def forward_from_state(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state)

    def value_from_state(self, state: torch.Tensor) -> torch.Tensor:
        state = self.head_dropout(state)
        return self.value_head(state)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        state, attention = self.forward_state(x)
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state), attention

    def forward_policy_only(self, x: torch.Tensor) -> torch.Tensor:
        state, _ = self.forward_state(x)
        state = self.head_dropout(state)
        return self.policy_head(state)

    def neighborhood_value(self, value_next: torch.Tensor) -> torch.Tensor:
        return self.omega_conv(value_next)

class SupervisedSMPModel(nn.Module):
    def __init__(
        self,
        *,
        arch: str,
        encoder_name: str,
        encoder_weights: str | None,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.smp_model = smp.create_model(
            arch=arch,
            encoder_name=encoder_name,
            encoder_weights=encoder_weights,
            in_channels=3,
            classes=1,
        )
        self.dropout = nn.Dropout2d(p=dropout_p)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        encoder_features = self.smp_model.encoder(x)
        decoder_output = run_smp_decoder(self.smp_model.decoder, encoder_features)
        decoder_output = self.dropout(decoder_output)
        logits = self.smp_model.segmentation_head(decoder_output)
        if getattr(self.smp_model, "classification_head", None) is not None:
            _ = self.smp_model.classification_head(encoder_features[-1])
        return logits

class SupervisedVGGModel(nn.Module):
    def __init__(
        self,
        *,
        num_scales: int,
        dilation: int,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.encoder = CustomVGGEncoderWrapper(num_scales=num_scales, dilation=dilation)
        self.segmentation_head = VGGSegmentationHead(
            pyramid_channels=self.encoder.pyramid_channels,
            dropout_p=dropout_p,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.segmentation_head(self.encoder.forward_pyramid(x))

class RefinementPolicyHead(nn.Module):
    def __init__(self, in_channels: int) -> None:
        super().__init__()
        self.body = nn.Sequential(
            _ConvBlock(in_channels, 512, dilation=1),
            _ConvBlock(512, 256, dilation=2),
            _ConvBlock(256, 128, dilation=3),
            _ConvBlock(128, 64, dilation=4),
        )
        self.classifier = nn.Conv2d(64, NUM_ACTIONS, kernel_size=1)
        nn.init.zeros_(self.classifier.weight)
        keep_bias = float(_job_param("keep_bias", 0.25))
        keep_index = NUM_ACTIONS // 2
        with torch.no_grad():
            self.classifier.bias.zero_()
            self.classifier.bias[keep_index] = keep_bias

class PixelDRLMG_WithDecoder(nn.Module):
    def __init__(
        self,
        *,
        arch: str,
        encoder_name: str,
        encoder_weights: str | None,
        encoder_depth: int,
        proj_dim: int,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.smp_model = smp.create_model(
            arch=arch,
            encoder_name=encoder_name,
            encoder_weights=encoder_weights,
            in_channels=3,
            classes=1,
        )
        self.smp_encoder = self.smp_model.encoder
        if REPLACE_BN_WITH_GN:
            replace_bn_with_gn(self.smp_encoder, num_groups=GN_NUM_GROUPS)

        raw_channels = sum(self.smp_encoder.out_channels)
        if proj_dim > 0:
            groups = min(GN_NUM_GROUPS, proj_dim)
            while groups > 1 and proj_dim % groups != 0:
                groups -= 1
            self.projection = nn.Sequential(
                nn.Conv2d(raw_channels, proj_dim, kernel_size=1, bias=False),
                nn.GroupNorm(groups, proj_dim) if REPLACE_BN_WITH_GN else nn.BatchNorm2d(proj_dim),
                nn.ReLU(inplace=True),
            )
            ch = proj_dim
        else:
            self.projection = None
            ch = raw_channels

        self.refinement_adapter = nn.Sequential(
            nn.Conv2d(ch + 6, ch, kernel_size=3, stride=1, padding=1, bias=False),
            _group_norm(ch),
            nn.ReLU(inplace=True),
            _ConvBlock(ch, ch, num_groups=GN_NUM_GROUPS),
        )
        self.sam = SelfAttentionModule(channels=ch)
        self.policy_head = RefinementPolicyHead(in_channels=ch)
        self.value_head = DilatedValueHead(in_channels=ch)
        self.head_dropout = nn.Dropout2d(p=dropout_p)
        self.omega_conv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
        nn.init.constant_(self.omega_conv.weight, 1.0 / 9.0)
        self.use_refinement = True

    def set_refinement_mode(self, enabled: bool) -> None:
        self.use_refinement = bool(enabled)
        self.refinement_adapter.requires_grad_(self.use_refinement)

    def _concat_from_features(
        self,
        features: list[torch.Tensor] | tuple[torch.Tensor, ...],
        *,
        output_size: tuple[int, int],
    ) -> torch.Tensor:
        h, w = output_size
        upsampled = []
        for feat in features:
            if feat.shape[-2] != h or feat.shape[-1] != w:
                feat = F.interpolate(feat, size=(h, w), mode="bilinear", align_corners=False)
            upsampled.append(feat)
        out = torch.cat(upsampled, dim=1)
        if self.projection is not None:
            out = self.projection(out)
        return out

    def _encoder_concat(self, x: torch.Tensor) -> torch.Tensor:
        return self._concat_from_features(self.smp_encoder(x), output_size=x.shape[-2:])

    def forward_decoder_from_features(
        self,
        encoder_features: list[torch.Tensor] | tuple[torch.Tensor, ...],
    ) -> torch.Tensor:
        decoder_output = run_smp_decoder(self.smp_model.decoder, encoder_features)
        logits = self.smp_model.segmentation_head(decoder_output)
        if getattr(self.smp_model, "classification_head", None) is not None:
            _ = self.smp_model.classification_head(encoder_features[-1])
        return logits

    def forward_decoder(self, x: torch.Tensor) -> torch.Tensor:
        return self.smp_model(x)

    def prepare_refinement_context(self, x: torch.Tensor) -> dict[str, torch.Tensor]:
        encoder_features = self.smp_encoder(x)
        decoder_logits = self.forward_decoder_from_features(encoder_features)
        return {
            "base_features": self._concat_from_features(encoder_features, output_size=x.shape[-2:]),
            "decoder_logits": decoder_logits,
            "decoder_prob": torch.sigmoid(decoder_logits),
        }

    def forward_refinement_state(
        self,
        base_features: torch.Tensor,
        current_mask: torch.Tensor,
        decoder_prob: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        residual = decoder_prob - current_mask
        uncertainty = (decoder_prob * (1.0 - decoder_prob)).clamp_(0.0, 0.25) * 4.0
        boundary = (current_mask - F.avg_pool2d(current_mask, kernel_size=3, stride=1, padding=1)).abs()
        conditioning = torch.cat(
            [decoder_prob, current_mask, residual, residual.abs(), uncertainty, boundary],
            dim=1,
        )
        fused = self.refinement_adapter(torch.cat([base_features, conditioning], dim=1))
        return self.sam(fused)

    def forward_state(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        concat_feat = self._encoder_concat(x)
        return self.sam(concat_feat)

    def forward_from_state(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state)

    def value_from_state(self, state: torch.Tensor) -> torch.Tensor:
        state = self.head_dropout(state)
        return self.value_head(state)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        if not self.use_refinement:
            state, attention = self.forward_state(x)
            return self.policy_head(state), self.value_head(state), attention
        context = self.prepare_refinement_context(x)
        state, attention = self.forward_refinement_state(
            context["base_features"],
            context["decoder_prob"].detach(),
            context["decoder_prob"],
        )
        policy_logits, value = self.forward_from_state(state)
        return policy_logits, value, attention

    def forward_policy_only(self, x: torch.Tensor) -> torch.Tensor:
        if not self.use_refinement:
            state, _ = self.forward_state(x)
            state = self.head_dropout(state)
            return self.policy_head(state)
        context = self.prepare_refinement_context(x)
        state, _ = self.forward_refinement_state(
            context["base_features"],
            context["decoder_prob"].detach(),
            context["decoder_prob"],
        )
        state = self.head_dropout(state)
        return self.policy_head(state)

    def neighborhood_value(self, value_next: torch.Tensor) -> torch.Tensor:
        return self.omega_conv(value_next)

class PixelDRLMG_VGGWithDecoder(nn.Module):
    def __init__(
        self,
        *,
        num_scales: int,
        dilation: int,
        dropout_p: float,
    ) -> None:
        super().__init__()
        self.encoder = CustomVGGEncoderWrapper(num_scales=num_scales, dilation=dilation)
        self.segmentation_head = VGGSegmentationHead(
            pyramid_channels=self.encoder.pyramid_channels,
            dropout_p=dropout_p,
        )
        ch = self.encoder.out_channels
        self.refinement_adapter = nn.Sequential(
            nn.Conv2d(ch + 3, ch, kernel_size=3, stride=1, padding=1, bias=False),
            _group_norm(ch),
            nn.ReLU(inplace=True),
            _ConvBlock(ch, ch, num_groups=GN_NUM_GROUPS),
        )
        self.sam = SelfAttentionModule(channels=ch)
        self.policy_head = DilatedPolicyHead(in_channels=ch)
        self.value_head = DilatedValueHead(in_channels=ch)
        self.head_dropout = nn.Dropout2d(p=dropout_p)
        self.omega_conv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)
        nn.init.constant_(self.omega_conv.weight, 1.0 / 9.0)
        self.use_refinement = True

    def set_refinement_mode(self, enabled: bool) -> None:
        self.use_refinement = bool(enabled)
        self.refinement_adapter.requires_grad_(self.use_refinement)

    def _concat_pyramid(
        self,
        pyramid: list[torch.Tensor] | tuple[torch.Tensor, ...],
        *,
        output_size: tuple[int, int],
    ) -> torch.Tensor:
        h, w = output_size
        upsampled = []
        for feat in pyramid:
            if feat.shape[-2] != h or feat.shape[-1] != w:
                feat = F.interpolate(feat, size=(h, w), mode="bilinear", align_corners=False)
            upsampled.append(feat)
        return torch.cat(upsampled, dim=1)

    def forward_decoder(self, x: torch.Tensor) -> torch.Tensor:
        return self.segmentation_head(self.encoder.forward_pyramid(x))

    def prepare_refinement_context(self, x: torch.Tensor) -> dict[str, torch.Tensor]:
        pyramid = self.encoder.forward_pyramid(x)
        decoder_logits = self.segmentation_head(pyramid)
        return {
            "base_features": self._concat_pyramid(pyramid, output_size=x.shape[-2:]),
            "decoder_logits": decoder_logits,
            "decoder_prob": torch.sigmoid(decoder_logits),
        }

    def forward_refinement_state(
        self,
        base_features: torch.Tensor,
        current_mask: torch.Tensor,
        decoder_prob: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        conditioning = torch.cat([decoder_prob, current_mask, decoder_prob - current_mask], dim=1)
        fused = self.refinement_adapter(torch.cat([base_features, conditioning], dim=1))
        return self.sam(fused)

    def forward_state(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        return self.sam(self.encoder(x))

    def forward_from_state(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        state = self.head_dropout(state)
        return self.policy_head(state), self.value_head(state)

    def value_from_state(self, state: torch.Tensor) -> torch.Tensor:
        state = self.head_dropout(state)
        return self.value_head(state)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        if not self.use_refinement:
            state, attention = self.forward_state(x)
            return self.policy_head(state), self.value_head(state), attention
        context = self.prepare_refinement_context(x)
        state, attention = self.forward_refinement_state(
            context["base_features"],
            context["decoder_prob"].detach(),
            context["decoder_prob"],
        )
        policy_logits, value = self.forward_from_state(state)
        return policy_logits, value, attention

    def forward_policy_only(self, x: torch.Tensor) -> torch.Tensor:
        if not self.use_refinement:
            state, _ = self.forward_state(x)
            state = self.head_dropout(state)
            return self.policy_head(state)
        context = self.prepare_refinement_context(x)
        state, _ = self.forward_refinement_state(
            context["base_features"],
            context["decoder_prob"].detach(),
            context["decoder_prob"],
        )
        state = self.head_dropout(state)
        return self.policy_head(state)

    def neighborhood_value(self, value_next: torch.Tensor) -> torch.Tensor:
        return self.omega_conv(value_next)

def run_smp_decoder(decoder: nn.Module, encoder_features: list[torch.Tensor] | tuple[torch.Tensor, ...]) -> torch.Tensor:
    signature = inspect.signature(decoder.forward)
    parameters = list(signature.parameters.values())
    if any(param.kind == inspect.Parameter.VAR_POSITIONAL for param in parameters):
        return decoder(*encoder_features)
    if len(parameters) == 1:
        return decoder(encoder_features)
    return decoder(*encoder_features)

def checkpoint_run_config_payload(payload: dict[str, Any]) -> dict[str, Any]:
    return payload.get("run_config") or payload.get("config") or {}

def _raw_decoder_rl_model(
    model: nn.Module,
) -> PixelDRLMG_WithDecoder | PixelDRLMG_VGGWithDecoder | None:
    raw = _unwrap_compiled(model)
    if isinstance(raw, (PixelDRLMG_WithDecoder, PixelDRLMG_VGGWithDecoder)):
        return raw
    return None

def _uses_refinement_runtime(model: nn.Module, *, strategy: int | None = None) -> bool:
    raw = _raw_decoder_rl_model(model)
    if raw is None:
        return False
    if strategy is not None and strategy not in (3, 4):
        return False
    return bool(getattr(raw, "use_refinement", False))

def _policy_action_count_from_state_dict(state_dict: dict[str, Any]) -> int | None:
    for key in (
        "policy_head.classifier.weight",
        "policy_head.classifier.bias",
        "policy_head.net.4.weight",
        "policy_head.net.4.bias",
    ):
        tensor = state_dict.get(key)
        if torch.is_tensor(tensor):
            return int(tensor.shape[0])
    return None

def _model_policy_action_count(model: nn.Module) -> int | None:
    raw = _unwrap_compiled(model)
    classifier = getattr(getattr(raw, "policy_head", None), "classifier", None)
    if isinstance(classifier, nn.Conv2d):
        return int(classifier.out_channels)
    return None

def _set_model_policy_action_count(model: nn.Module, action_count: int) -> bool:
    raw = _unwrap_compiled(model)
    policy_head = getattr(raw, "policy_head", None)
    classifier = getattr(policy_head, "classifier", None)
    if not isinstance(classifier, nn.Conv2d):
        return False
    if int(classifier.out_channels) == int(action_count):
        return False

    new_classifier = nn.Conv2d(
        classifier.in_channels,
        int(action_count),
        kernel_size=classifier.kernel_size,
        stride=classifier.stride,
        padding=classifier.padding,
        dilation=classifier.dilation,
        groups=classifier.groups,
        bias=classifier.bias is not None,
        padding_mode=classifier.padding_mode,
    ).to(device=classifier.weight.device, dtype=classifier.weight.dtype)
    nn.init.zeros_(new_classifier.weight)
    if new_classifier.bias is not None:
        bias = torch.full((int(action_count),), -2.0, dtype=new_classifier.bias.dtype, device=new_classifier.bias.device)
        keep_index = 1 if int(action_count) >= 3 else int(action_count) - 1
        bias[keep_index] = 2.0
        with torch.no_grad():
            new_classifier.bias.copy_(bias)
    policy_head.classifier = new_classifier
    return True

def _configure_policy_head_compatibility(
    model: nn.Module,
    state_dict: dict[str, Any],
    *,
    source: str,
) -> int | None:
    action_count = _policy_action_count_from_state_dict(state_dict)
    if action_count is None:
        return None
    if _set_model_policy_action_count(model, action_count):
        print(f"[Policy Compatibility] source={source} num_actions={action_count}")
    return action_count

def _strategy34_checkpoint_layout_info(
    state_dict: dict[str, Any],
    run_config: dict[str, Any] | None = None,
) -> dict[str, Any]:
    run_config = run_config or {}
    strategy = run_config.get("strategy")
    has_legacy_policy_head = any(key.startswith("policy_head.net.") for key in state_dict)
    has_new_policy_body = any(key.startswith("policy_head.body.") for key in state_dict)
    has_new_policy_classifier = any(key.startswith("policy_head.classifier.") for key in state_dict)
    has_refinement_adapter = any(key.startswith("refinement_adapter.") for key in state_dict)
    is_strategy34_decoder_checkpoint = bool(
        strategy in (3, 4)
        or has_legacy_policy_head
        or has_new_policy_body
        or has_new_policy_classifier
        or has_refinement_adapter
    )
    use_refinement = bool(
        has_refinement_adapter or ((has_new_policy_body or has_new_policy_classifier) and not has_legacy_policy_head)
    )
    return {
        "strategy": strategy,
        "is_strategy34_decoder_checkpoint": is_strategy34_decoder_checkpoint,
        "has_legacy_policy_head": has_legacy_policy_head,
        "has_new_policy_head": bool(has_new_policy_body or has_new_policy_classifier),
        "has_refinement_adapter": has_refinement_adapter,
        "requires_policy_remap": has_legacy_policy_head,
        "policy_action_count": _policy_action_count_from_state_dict(state_dict),
        "use_refinement": use_refinement,
        "compatibility_mode": "refinement" if use_refinement else "legacy",
    }

def inspect_strategy34_checkpoint_compatibility(path: str | Path) -> dict[str, Any]:
    checkpoint_path = Path(path).expanduser().resolve()
    payload = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
    layout = _strategy34_checkpoint_layout_info(payload.get("model_state_dict", {}), checkpoint_run_config_payload(payload))
    layout["path"] = str(checkpoint_path)
    return layout

def _configure_strategy34_model_compatibility(
    model: nn.Module,
    layout: dict[str, Any],
    *,
    source: str,
) -> None:
    raw = _raw_decoder_rl_model(model)
    if raw is None or not layout.get("is_strategy34_decoder_checkpoint"):
        return
    raw.set_refinement_mode(bool(layout["use_refinement"]))
    if not bool(layout["use_refinement"]):
        raw.refinement_adapter.eval()
    print(
        "[Strategy34 Compatibility] "
        f"source={source} mode={layout['compatibility_mode']} "
        f"legacy_policy_head={layout['has_legacy_policy_head']} "
        f"refinement_adapter={layout['has_refinement_adapter']}"
    )

def _configure_model_from_checkpoint_path(
    model: nn.Module,
    checkpoint_path: str | Path,
) -> dict[str, Any]:
    checkpoint_path = Path(checkpoint_path).expanduser().resolve()
    payload = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
    state_dict = payload.get("model_state_dict", {})
    _configure_policy_head_compatibility(model, state_dict, source=str(checkpoint_path))
    layout = _strategy34_checkpoint_layout_info(state_dict, checkpoint_run_config_payload(payload))
    _configure_strategy34_model_compatibility(model, layout, source=str(checkpoint_path))
    layout["path"] = str(checkpoint_path)
    return layout

def _remap_legacy_policy_head_state_dict(state_dict: dict[str, Any]) -> dict[str, Any]:
    remapped: dict[str, Any] = {}
    for key, value in state_dict.items():
        if key.startswith("policy_head.net."):
            suffix = key[len("policy_head.net."):]
            layer_idx, dot, rest = suffix.partition(".")
            if dot:
                if layer_idx in {"0", "1", "2", "3"}:
                    remapped[f"policy_head.body.{layer_idx}.{rest}"] = value
                    continue
                if layer_idx == "4":
                    remapped[f"policy_head.classifier.{rest}"] = value
                    continue
        remapped[key] = value
    return remapped

def _load_strategy2_checkpoint_payload(
    path: str | Path,
    *,
    model_config: RuntimeModelConfig,
) -> dict[str, Any]:
    checkpoint_path = Path(path).expanduser().resolve()
    ckpt = torch.load(checkpoint_path, map_location=DEVICE, weights_only=False)
    saved_config = checkpoint_run_config_payload(ckpt)
    if saved_config:
        saved_model_config = RuntimeModelConfig.from_payload(saved_config).validate()
        if saved_model_config.backbone_family != model_config.backbone_family:
            raise ValueError(
                f"Strategy 2 checkpoint backbone family mismatch: requested {model_config.backbone_family!r}, "
                f"checkpoint has {saved_model_config.backbone_family!r} at {checkpoint_path}."
            )
    return ckpt

def _use_channels_last_for_run(model_config: RuntimeModelConfig | None = None) -> bool:
    model_config = (model_config or current_model_config()).validate()
    if not USE_CHANNELS_LAST:
        return False
    if DEVICE.type != "cuda":
        return False
    amp_dtype = resolve_amp_dtype(AMP_DTYPE)
    if (
        model_config.backbone_family == "smp"
        and "efficientnet" in model_config.smp_encoder_name.lower()
        and USE_AMP
        and amp_dtype in {torch.float16, torch.bfloat16}
    ):
        print("[MemoryFormat] Disabling channels_last for EfficientNet + AMP stability.")
        return False
    return True

def build_model(
    strategy: int,
    dropout_p: float,
    *,
    model_config: RuntimeModelConfig,
    strategy2_checkpoint_path: str | Path | None = None,
) -> tuple[nn.Module, str, bool]:
    model_config = model_config.validate()
    if model_config.backbone_family == "custom_vgg":
        if strategy == 1:
            model = PixelDRLMG_VGG(
                num_scales=model_config.vgg_feature_scales,
                dilation=model_config.vgg_feature_dilation,
                dropout_p=dropout_p,
            )
        elif strategy == 2:
            model = SupervisedVGGModel(
                num_scales=model_config.vgg_feature_scales,
                dilation=model_config.vgg_feature_dilation,
                dropout_p=dropout_p,
            )
        elif strategy == 3:
            model = PixelDRLMG_VGGWithDecoder(
                num_scales=model_config.vgg_feature_scales,
                dilation=model_config.vgg_feature_dilation,
                dropout_p=dropout_p,
            )
            if strategy2_checkpoint_path is not None:
                ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
                s2_state = ckpt["model_state_dict"]
                decoder_state = {
                    key: value
                    for key, value in s2_state.items()
                    if key.startswith("encoder.") or key.startswith("segmentation_head.")
                }
                model.load_state_dict(decoder_state, strict=False)
        elif strategy == 4:
            if strategy2_checkpoint_path is None:
                raise ValueError("Strategy 4 requires a Strategy 2 checkpoint path.")
            model = PixelDRLMG_VGGWithDecoder(
                num_scales=model_config.vgg_feature_scales,
                dilation=model_config.vgg_feature_dilation,
                dropout_p=dropout_p,
            )
            ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
            s2_state = ckpt["model_state_dict"]
            decoder_state = {
                key: value
                for key, value in s2_state.items()
                if key.startswith("encoder.") or key.startswith("segmentation_head.")
            }
            model.load_state_dict(decoder_state, strict=False)
            model.encoder.requires_grad_(False)
            model.segmentation_head.requires_grad_(False)
        elif strategy == 5:
            if strategy2_checkpoint_path is None:
                raise ValueError("Strategy 5 requires a Strategy 2 checkpoint path.")
            model = PixelDRLMG_VGG(
                num_scales=model_config.vgg_feature_scales,
                dilation=model_config.vgg_feature_dilation,
                dropout_p=dropout_p,
            )
            ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
            s2_state = ckpt["model_state_dict"]
            encoder_state = {}
            for key, value in s2_state.items():
                if key.startswith("encoder."):
                    encoder_state[key.replace("encoder.", "extractor.encoder.")] = value
            model.load_state_dict(encoder_state, strict=False)
            model.extractor.encoder.requires_grad_(False)
        else:
            raise ValueError(f"Unsupported strategy: {strategy}")
    else:
        if strategy == 1:
            model = PixelDRLMG_SMP(
                encoder_name=model_config.smp_encoder_name,
                encoder_weights=model_config.smp_encoder_weights,
                encoder_depth=model_config.smp_encoder_depth,
                proj_dim=model_config.smp_encoder_proj_dim,
                dropout_p=dropout_p,
            )
        elif strategy == 2:
            model = SupervisedSMPModel(
                arch=model_config.smp_decoder_type,
                encoder_name=model_config.smp_encoder_name,
                encoder_weights=model_config.smp_encoder_weights,
                dropout_p=dropout_p,
            )
        elif strategy == 3:
            model = PixelDRLMG_WithDecoder(
                arch=model_config.smp_decoder_type,
                encoder_name=model_config.smp_encoder_name,
                encoder_weights=model_config.smp_encoder_weights,
                encoder_depth=model_config.smp_encoder_depth,
                proj_dim=model_config.smp_encoder_proj_dim,
                dropout_p=dropout_p,
            )
            if strategy2_checkpoint_path is not None:
                ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
                s2_state = ckpt["model_state_dict"]
                smp_state = {k: v for k, v in s2_state.items() if k.startswith("smp_model.")}
                model.load_state_dict(smp_state, strict=False)
        elif strategy == 4:
            if strategy2_checkpoint_path is None:
                raise ValueError("Strategy 4 requires a Strategy 2 checkpoint path.")
            model = PixelDRLMG_WithDecoder(
                arch=model_config.smp_decoder_type,
                encoder_name=model_config.smp_encoder_name,
                encoder_weights=None,
                encoder_depth=model_config.smp_encoder_depth,
                proj_dim=model_config.smp_encoder_proj_dim,
                dropout_p=dropout_p,
            )
            ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
            s2_state = ckpt["model_state_dict"]
            smp_state = {k: v for k, v in s2_state.items() if k.startswith("smp_model.")}
            model.load_state_dict(smp_state, strict=False)
            model.smp_model.encoder.requires_grad_(False)
            model.smp_model.decoder.requires_grad_(False)
            if hasattr(model.smp_model, "segmentation_head"):
                model.smp_model.segmentation_head.requires_grad_(False)
        elif strategy == 5:
            if strategy2_checkpoint_path is None:
                raise ValueError("Strategy 5 requires a Strategy 2 checkpoint path.")
            model = PixelDRLMG_SMP(
                encoder_name=model_config.smp_encoder_name,
                encoder_weights=None,
                encoder_depth=model_config.smp_encoder_depth,
                proj_dim=model_config.smp_encoder_proj_dim,
                dropout_p=dropout_p,
            )
            ckpt = _load_strategy2_checkpoint_payload(strategy2_checkpoint_path, model_config=model_config)
            s2_state = ckpt["model_state_dict"]
            encoder_state = {}
            for key, value in s2_state.items():
                if key.startswith("smp_model.encoder."):
                    encoder_state[key.replace("smp_model.encoder.", "extractor.encoder.")] = value
            model.load_state_dict(encoder_state, strict=False)
            model.extractor.encoder.requires_grad_(False)
        else:
            raise ValueError(f"Unsupported strategy: {strategy}")

    use_channels_last_now = _use_channels_last_for_run(model_config)
    model = model.to(DEVICE)
    if use_channels_last_now:
        model = model.to(memory_format=torch.channels_last)

    compiled = False
    if USE_TORCH_COMPILE and hasattr(torch, "compile"):
        try:
            model = torch.compile(model)
            compiled = True
        except Exception as exc:
            print(f"[Compile] torch.compile skipped: {exc}")
    return model, strategy_name(strategy, model_config), compiled

def _unwrap_compiled(model: nn.Module) -> nn.Module:
    return getattr(model, "_orig_mod", model)

def count_parameters(module: nn.Module | None, *, only_trainable: bool = False) -> int:
    if module is None:
        return 0
    if only_trainable:
        return sum(p.numel() for p in module.parameters() if p.requires_grad)
    return sum(p.numel() for p in module.parameters())

def print_model_parameter_summary(
    *,
    model: nn.Module,
    description: str,
    strategy: int,
    model_config: RuntimeModelConfig,
    dropout_p: float,
    amp_dtype: torch.dtype,
    compiled: bool,
) -> None:
    raw = _unwrap_compiled(model)
    total_params = count_parameters(raw)
    trainable_params = count_parameters(raw, only_trainable=True)
    frozen_params = total_params - trainable_params
    bn_count = sum(1 for m in raw.modules() if isinstance(m, (nn.BatchNorm2d, nn.BatchNorm1d)))
    gn_count = sum(1 for m in raw.modules() if isinstance(m, nn.GroupNorm))

    section(f"Model Parameter Summary | {description}")
    print(f"Strategy              : {strategy}")
    print(f"Model                 : {description}")
    print(f"Dropout p             : {dropout_p:.4f}")
    print(f"Total params          : {total_params:,}")
    print(f"Trainable params      : {trainable_params:,}")
    print(f"Frozen params         : {frozen_params:,}")
    print(f"BN layers             : {bn_count}")
    print(f"GN layers             : {gn_count}")
    print(f"channels_last         : {_use_channels_last_for_run(model_config)}")
    print(f"AMP dtype             : {amp_dtype}")
    print(f"torch.compile         : {compiled}")
    print(f"Backbone family       : {model_config.backbone_family}")

    block_counts: dict[str, int] = {}
    if strategy == 1:
        block_counts["encoder"] = count_parameters(getattr(raw.extractor, "encoder", None))
        block_counts["projection"] = count_parameters(getattr(raw.extractor, "projection", None))
        block_counts["sam"] = count_parameters(getattr(raw, "sam", None))
        block_counts["policy_head"] = count_parameters(getattr(raw, "policy_head", None))
        block_counts["value_head"] = count_parameters(getattr(raw, "value_head", None))
        block_counts["omega"] = count_parameters(getattr(raw, "omega_conv", None))
    elif strategy == 2:
        if hasattr(raw, "smp_model"):
            smp_model = raw.smp_model
            block_counts["encoder"] = count_parameters(getattr(smp_model, "encoder", None))
            block_counts["decoder"] = count_parameters(getattr(smp_model, "decoder", None))
            block_counts["segmentation_head"] = count_parameters(getattr(smp_model, "segmentation_head", None))
            block_counts["dropout"] = count_parameters(getattr(raw, "dropout", None))
        else:
            block_counts["encoder"] = count_parameters(getattr(raw, "encoder", None))
            block_counts["segmentation_head"] = count_parameters(getattr(raw, "segmentation_head", None))
    elif strategy in (3, 4):
        if hasattr(raw, "smp_model"):
            smp_model = raw.smp_model
            block_counts["encoder"] = count_parameters(getattr(smp_model, "encoder", None))
            block_counts["decoder"] = count_parameters(getattr(smp_model, "decoder", None))
            block_counts["segmentation_head"] = count_parameters(getattr(smp_model, "segmentation_head", None))
        else:
            block_counts["encoder"] = count_parameters(getattr(raw, "encoder", None))
            block_counts["segmentation_head"] = count_parameters(getattr(raw, "segmentation_head", None))
        block_counts["sam"] = count_parameters(getattr(raw, "sam", None))
        block_counts["policy_head"] = count_parameters(getattr(raw, "policy_head", None))
        block_counts["value_head"] = count_parameters(getattr(raw, "value_head", None))
        block_counts["omega"] = count_parameters(getattr(raw, "omega_conv", None))
    elif strategy == 5:
        block_counts["encoder"] = count_parameters(getattr(raw.extractor, "encoder", None))
        block_counts["projection"] = count_parameters(getattr(raw.extractor, "projection", None))
        block_counts["sam"] = count_parameters(getattr(raw, "sam", None))
        block_counts["policy_head"] = count_parameters(getattr(raw, "policy_head", None))
        block_counts["value_head"] = count_parameters(getattr(raw, "value_head", None))
        block_counts["omega"] = count_parameters(getattr(raw, "omega_conv", None))

    for name, value in block_counts.items():
        print(f"{name:22s}: {value:,}")

"""=============================================================================
METRICS + CHECKPOINTS
=============================================================================
"""

_EPS = 1e-4

def _as_bool(mask: np.ndarray) -> np.ndarray:
    return (mask[0] if mask.ndim == 3 else mask).astype(bool)

def _tp_fp_fn(pred: np.ndarray, target: np.ndarray):
    p, t = _as_bool(pred), _as_bool(target)
    tp = float((p & t).sum())
    fp = float((p & ~t).sum())
    fn = float((~p & t).sum())
    return tp, fp, fn, float(t.sum()), float(p.sum())

def dice_score(pred: np.ndarray, target: np.ndarray) -> float:
    tp, _, _, t, p = _tp_fp_fn(pred, target)
    return (2 * tp + _EPS) / (t + p + _EPS)

def ppv_score(pred: np.ndarray, target: np.ndarray) -> float:
    tp, fp, *_ = _tp_fp_fn(pred, target)
    return (tp + _EPS) / (tp + fp + _EPS)

def sensitivity_score(pred: np.ndarray, target: np.ndarray) -> float:
    tp, _, fn, *_ = _tp_fp_fn(pred, target)
    return (tp + _EPS) / (tp + fn + _EPS)

def iou_score(pred: np.ndarray, target: np.ndarray) -> float:
    tp, _, _, t, p = _tp_fp_fn(pred, target)
    return (tp + _EPS) / (t + p - tp + _EPS)

def _boundary(mask: np.ndarray) -> np.ndarray:
    m = _as_bool(mask)
    if not m.any():
        return m
    return m ^ ndimage.binary_erosion(m, iterations=1, border_value=0)

def boundary_iou_score(pred: np.ndarray, target: np.ndarray) -> float:
    pb, tb = _boundary(pred), _boundary(target)
    inter = float((pb & tb).sum())
    union = float((pb | tb).sum())
    return (inter + _EPS) / (union + _EPS)

def _surf_dist(a: np.ndarray, b: np.ndarray) -> np.ndarray:
    a, b = _as_bool(a), _as_bool(b)
    if not a.any() and not b.any():
        return np.array([0.0], dtype=np.float32)
    if not a.any() or not b.any():
        return np.array([np.inf], dtype=np.float32)
    ba, bb = _boundary(a), _boundary(b)
    return ndimage.distance_transform_edt(~bb)[ba].astype(np.float32)

def hd95_score(pred: np.ndarray, target: np.ndarray) -> float:
    distances = np.concatenate([_surf_dist(pred, target), _surf_dist(target, pred)])
    return float("inf") if np.isinf(distances).any() else float(np.percentile(distances, 95))

def compute_all_metrics(pred: np.ndarray, target: np.ndarray) -> dict[str, float]:
    return {
        "dice": dice_score(pred, target),
        "ppv": ppv_score(pred, target),
        "sen": sensitivity_score(pred, target),
        "iou": iou_score(pred, target),
        "biou": boundary_iou_score(pred, target),
        "hd95": hd95_score(pred, target),
    }

def checkpoint_manifest_path(path: Path) -> Path:
    path = Path(path)
    return path.with_name(f"{path.name}.meta.json")

def checkpoint_history_path(run_dir: Path, run_type: str) -> Path:
    if run_type == "overfit":
        return Path(run_dir) / "overfit_history.json"
    return Path(run_dir) / "history.json"

def checkpoint_state_presence(payload: dict[str, Any]) -> dict[str, bool]:
    tracked = [
        "model_state_dict",
        "optimizer_state_dict",
        "scheduler_state_dict",
        "scaler_state_dict",
        "log_alpha",
        "alpha_optimizer_state_dict",
        "best_metric_name",
        "best_metric_value",
        "patience_counter",
        "elapsed_seconds",
        "run_config",
        "epoch_metrics",
        "resume_source",
    ]
    return {name: name in payload for name in tracked}

def write_checkpoint_manifest(
    path: Path,
    payload: dict[str, Any],
    *,
    extra: dict[str, Any] | None = None,
) -> dict[str, Any]:
    run_config = checkpoint_run_config_payload(payload)
    manifest = {
        "checkpoint_path": str(Path(path).resolve()),
        "run_type": payload.get("run_type", "unknown"),
        "epoch": int(payload.get("epoch", 0)),
        "strategy": run_config.get("strategy"),
        "dataset_percent": run_config.get("dataset_percent"),
        "backbone_family": run_config.get("backbone_family", "smp"),
        "saved_keys": sorted(payload.keys()),
        "state_presence": checkpoint_state_presence(payload),
    }
    if "resume_source" in payload:
        manifest["resume_source"] = payload["resume_source"]
    if extra:
        manifest.update(extra)
    save_json(checkpoint_manifest_path(path), manifest)
    return manifest

def checkpoint_required_keys(
    *,
    optimizer: torch.optim.Optimizer | None,
    scheduler: CosineAnnealingLR | None,
    scaler: Any | None,
    log_alpha: torch.Tensor | None,
    alpha_optimizer: torch.optim.Optimizer | None,
    require_run_metadata: bool,
) -> list[str]:
    keys = ["epoch", "model_state_dict"]
    if require_run_metadata:
        keys.extend(
            [
                "run_type",
                "best_metric_name",
                "best_metric_value",
                "patience_counter",
                "elapsed_seconds",
                "run_config",
                "epoch_metrics",
            ]
        )
    if optimizer is not None:
        keys.append("optimizer_state_dict")
    if scheduler is not None:
        keys.append("scheduler_state_dict")
    if scaler is not None:
        keys.append("scaler_state_dict")
    if log_alpha is not None:
        keys.append("log_alpha")
    if alpha_optimizer is not None:
        keys.append("alpha_optimizer_state_dict")
    return keys

def validate_checkpoint_payload(
    path: Path,
    payload: dict[str, Any],
    *,
    required_keys: list[str],
    expected_run_type: str | None = None,
) -> None:
    missing = [name for name in required_keys if name not in payload]
    if missing:
        raise KeyError(f"Checkpoint {path} is missing required keys: {missing}")
    if expected_run_type is not None and payload.get("run_type") != expected_run_type:
        raise ValueError(
            f"Checkpoint {path} run_type mismatch: expected {expected_run_type!r}, "
            f"got {payload.get('run_type')!r}."
        )

def save_checkpoint(
    path: Path,
    *,
    run_type: str,
    model: nn.Module,
    optimizer: torch.optim.Optimizer,
    scheduler: ReduceLROnPlateau | None,
    scaler: Any | None,
    epoch: int,
    best_metric_value: float,
    best_metric_name: str,
    run_config: dict[str, Any],
    epoch_metrics: dict[str, Any],
    patience_counter: int,
    elapsed_seconds: float,
    log_alpha: torch.Tensor | None = None,
    alpha_optimizer: torch.optim.Optimizer | None = None,
    resume_source: dict[str, Any] | None = None,
) -> None:
    payload = {
        "run_type": run_type,
        "epoch": epoch,
        "model_state_dict": _unwrap_compiled(model).state_dict(),
        "optimizer_state_dict": optimizer.state_dict(),
        "best_metric_name": best_metric_name,
        "best_metric_value": best_metric_value,
        "patience_counter": int(patience_counter),
        "elapsed_seconds": float(elapsed_seconds),
        "run_config": run_config,
        "config": run_config,
        "epoch_metrics": epoch_metrics,
    }
    if scheduler is not None:
        payload["scheduler_state_dict"] = scheduler.state_dict()
    if scaler is not None:
        payload["scaler_state_dict"] = scaler.state_dict()
    if log_alpha is not None:
        payload["log_alpha"] = float(log_alpha.detach().item())
    if alpha_optimizer is not None:
        payload["alpha_optimizer_state_dict"] = alpha_optimizer.state_dict()
    if resume_source is not None:
        payload["resume_source"] = resume_source
    validate_checkpoint_payload(
        path,
        payload,
        required_keys=checkpoint_required_keys(
            optimizer=optimizer,
            scheduler=scheduler,
            scaler=scaler,
            log_alpha=log_alpha,
            alpha_optimizer=alpha_optimizer,
            require_run_metadata=True,
        ),
        expected_run_type=run_type,
    )
    torch.save(payload, path)
    write_checkpoint_manifest(path, payload)

def load_checkpoint(
    path: Path,
    *,
    model: nn.Module,
    optimizer: torch.optim.Optimizer | None = None,
    scheduler: ReduceLROnPlateau | None = None,
    scaler: Any | None = None,
    device: torch.device,
    log_alpha: torch.Tensor | None = None,
    alpha_optimizer: torch.optim.Optimizer | None = None,
    expected_run_type: str | None = None,
    require_run_metadata: bool = False,
) -> dict[str, Any]:
    ckpt = torch.load(path, map_location=device, weights_only=False)
    validate_checkpoint_payload(
        path,
        ckpt,
        required_keys=checkpoint_required_keys(
            optimizer=optimizer,
            scheduler=scheduler,
            scaler=scaler,
            log_alpha=log_alpha,
            alpha_optimizer=alpha_optimizer,
            require_run_metadata=require_run_metadata,
        ),
        expected_run_type=expected_run_type,
    )

    raw_model = _unwrap_compiled(model)
    state_dict = ckpt["model_state_dict"]
    load_strict = True
    compat_layout: dict[str, Any] | None = None
    _configure_policy_head_compatibility(model, state_dict, source=str(path))
    if any(key.startswith("policy_head.net.") for key in state_dict):
        state_dict = _remap_legacy_policy_head_state_dict(state_dict)
    if isinstance(raw_model, (PixelDRLMG_WithDecoder, PixelDRLMG_VGGWithDecoder)):
        compat_layout = _strategy34_checkpoint_layout_info(ckpt["model_state_dict"], checkpoint_run_config_payload(ckpt))
        if compat_layout["is_strategy34_decoder_checkpoint"]:
            _configure_strategy34_model_compatibility(model, compat_layout, source=str(path))
            load_strict = bool(compat_layout["use_refinement"])

    incompatible = raw_model.load_state_dict(state_dict, strict=load_strict)
    if not load_strict:
        missing_keys = [key for key in incompatible.missing_keys if not key.startswith("refinement_adapter.")]
        unexpected_keys = list(incompatible.unexpected_keys)
        if missing_keys or unexpected_keys:
            print(
                "[Checkpoint Restore] Non-strict legacy Strategy 3/4 load "
                f"missing={missing_keys} unexpected={unexpected_keys}"
            )

    if optimizer is not None and "optimizer_state_dict" in ckpt:
        try:
            optimizer.load_state_dict(ckpt["optimizer_state_dict"])
        except ValueError:
            if compat_layout is None or compat_layout.get("compatibility_mode") != "legacy":
                raise
            print(
                f"[Checkpoint Restore] Skipping optimizer state for legacy Strategy 3/4 checkpoint at {path} "
                "because the parameter layout differs from the refinement-capable model."
            )
    if scheduler is not None and "scheduler_state_dict" in ckpt:
        scheduler.load_state_dict(ckpt["scheduler_state_dict"])
    if scaler is not None and "scaler_state_dict" in ckpt:
        scaler.load_state_dict(ckpt["scaler_state_dict"])
    if log_alpha is not None and "log_alpha" in ckpt:
        with torch.no_grad():
            log_alpha.fill_(float(ckpt["log_alpha"]))
    if alpha_optimizer is not None and "alpha_optimizer_state_dict" in ckpt:
        alpha_optimizer.load_state_dict(ckpt["alpha_optimizer_state_dict"])
    restored = checkpoint_state_presence(ckpt)
    restore_info = {
        "restored_keys": restored,
        "restored_at_epoch": int(ckpt.get("epoch", 0)),
        "expected_run_type": expected_run_type,
    }
    write_checkpoint_manifest(path, ckpt, extra={"last_restore": restore_info})
    print(
        f"[Checkpoint Restore] path={path} epoch={ckpt.get('epoch')} "
        f"run_type={ckpt.get('run_type', 'unknown')} "
        f"backbone={checkpoint_run_config_payload(ckpt).get('backbone_family', 'unknown')}"
    )
    return ckpt

"""=============================================================================
TRAINING + VALIDATION
=============================================================================
"""

def sample_actions(policy_logits: torch.Tensor, stochastic: bool) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    logits = policy_logits.float()
    log_probs = F.log_softmax(logits, dim=1)
    probs = log_probs.exp()
    entropy = -(probs * log_probs).sum(dim=1).mean()
    if stochastic:
        uniform = torch.rand_like(logits)
        gumbel = -torch.log(-torch.log(uniform + 1e-8) + 1e-8)
        actions = (logits + gumbel).argmax(dim=1)
    else:
        actions = logits.argmax(dim=1)
    log_prob = log_probs.gather(1, actions.unsqueeze(1))
    return actions, log_prob, entropy

def apply_actions(
    seg: torch.Tensor,
    actions: torch.Tensor,
    *,
    soft_update_step: float | None = None,
    num_actions: int | None = None,
) -> torch.Tensor:
    num_actions = int(num_actions if num_actions is not None else _job_param("num_actions", NUM_ACTIONS))
    if num_actions == 5:
        action_map = actions.long()
        if soft_update_step is not None:
            deltas = _refinement_deltas(device=seg.device, dtype=seg.dtype)
            delta = deltas[action_map].unsqueeze(1)
            return (seg + delta).clamp_(0.0, 1.0)
        return torch.where(
            action_map.unsqueeze(1) <= 1,
            torch.zeros_like(seg),
            torch.where(action_map.unsqueeze(1) >= 3, torch.ones_like(seg), seg),
        )
    action_map = actions.unsqueeze(1)
    if num_actions >= 3:
        if soft_update_step is not None:
            delta = torch.where(
                action_map == 0,
                torch.full_like(seg, -soft_update_step),
                torch.where(action_map == 2, torch.full_like(seg, soft_update_step), torch.zeros_like(seg)),
            )
            return (seg + delta).clamp_(0.0, 1.0)
        return torch.where(
            action_map == 0,
            torch.zeros_like(seg),
            torch.where(action_map == 2, torch.ones_like(seg), seg),
        )
    return seg * (action_map == 1).to(dtype=seg.dtype)

def _soft_dice_tensor(pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
    inter = (pred * target).sum(dim=(1, 2, 3))
    denom = pred.sum(dim=(1, 2, 3)) + target.sum(dim=(1, 2, 3))
    return (2.0 * inter + 1e-6) / (denom + 1e-6)

def _soft_boundary(mask: torch.Tensor) -> torch.Tensor:
    return (mask - F.avg_pool2d(mask, kernel_size=3, stride=1, padding=1)).abs()

def compute_refinement_reward(
    seg: torch.Tensor,
    seg_next: torch.Tensor,
    gt_mask: torch.Tensor,
    *,
    decoder_prior: torch.Tensor | None = None,
) -> torch.Tensor:
    seg_f = seg.float()
    seg_next_f = seg_next.float()
    gt_f = gt_mask.float()

    base_reward = (seg_f - gt_f).pow(2) - (seg_next_f - gt_f).pow(2)
    if decoder_prior is not None:
        prior_f = decoder_prior.float()
        uncertainty_focus = (4.0 * prior_f * (1.0 - prior_f)).clamp(0.0, 1.0)
        base_reward = base_reward * (0.35 + 0.65 * uncertainty_focus)
    reward = base_reward

    dice_weight = float(_job_param("dice_reward_weight", 0.35))
    if dice_weight > 0:
        prev_dice = _soft_dice_tensor(seg_f, gt_f)
        next_dice = _soft_dice_tensor(seg_next_f, gt_f)
        reward = reward + dice_weight * (next_dice - prev_dice).view(-1, 1, 1, 1)

    boundary_weight = float(_job_param("boundary_reward_weight", 0.15))
    if boundary_weight > 0:
        gt_boundary = _soft_boundary(gt_f)
        prev_boundary = _soft_boundary(seg_f)
        next_boundary = _soft_boundary(seg_next_f)
        boundary_reward = (prev_boundary - gt_boundary).pow(2) - (next_boundary - gt_boundary).pow(2)
        reward = reward + boundary_weight * boundary_reward

    prior_weight = float(_job_param("prior_reward_weight", 0.01))
    if decoder_prior is not None and prior_weight > 0:
        prior_f = decoder_prior.float()
        confidence = ((prior_f - 0.5).abs() * 2.0).clamp(0.0, 1.0)
        reward = reward + prior_weight * confidence * (
            (seg_f - prior_f).pow(2) - (seg_next_f - prior_f).pow(2)
        )
    return reward

def compute_strategy1_aux_segmentation_loss(
    policy_logits: torch.Tensor,
    gt_mask: torch.Tensor,
    *,
    ce_weight: float,
    dice_weight: float,
    seg_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, float, float]:
    if int(policy_logits.shape[1]) == 5:
        logits_f = policy_logits.float()
        gt_mask_f = gt_mask.float()
        base_seg = seg_mask.float() if seg_mask is not None else torch.full_like(gt_mask_f, 0.5)
        probs = F.softmax(logits_f, dim=1)
        deltas = _refinement_deltas(device=logits_f.device, dtype=logits_f.dtype).view(1, -1, 1, 1)
        expected_delta = (probs * deltas).sum(dim=1, keepdim=True)
        predicted_next = (base_seg + expected_delta).clamp(1e-4, 1.0 - 1e-4)

        aux_loss = torch.zeros((), device=policy_logits.device, dtype=torch.float32)
        ce_loss_value = 0.0
        dice_loss_value = 0.0

        if ce_weight > 0:
            bce = F.binary_cross_entropy(predicted_next, gt_mask_f)
            aux_loss = aux_loss + ce_weight * bce
            ce_loss_value = float(bce.detach().item())

        if dice_weight > 0:
            inter = (predicted_next * gt_mask_f).sum()
            dice_loss = 1.0 - (2.0 * inter + 1e-6) / (predicted_next.sum() + gt_mask_f.sum() + 1e-6)
            aux_loss = aux_loss + dice_weight * dice_loss
            dice_loss_value = float(dice_loss.detach().item())

        return aux_loss, ce_loss_value, dice_loss_value

    logits_f = policy_logits.float()
    gt_mask_f = gt_mask.float()
    num_actions = int(logits_f.shape[1])
    aux_loss = torch.zeros((), device=policy_logits.device, dtype=torch.float32)
    ce_loss_value = 0.0
    dice_loss_value = 0.0

    if ce_weight > 0:
        if num_actions >= 3:
            if seg_mask is None:
                seg_mask = torch.ones_like(gt_mask)
            seg_bin = threshold_binary_long(seg_mask).squeeze(1)
            gt_bin = gt_mask[:, 0].long()
            aux_target = torch.where(
                gt_bin == 0,
                torch.zeros_like(gt_bin),
                torch.where(seg_bin == 1, torch.ones_like(gt_bin), torch.full_like(gt_bin, 2)),
            )
        else:
            aux_target = gt_mask[:, 0].long() * (num_actions - 1)
        ce_loss = F.cross_entropy(logits_f, aux_target)
        aux_loss = aux_loss + ce_weight * ce_loss
        ce_loss_value = float(ce_loss.detach().item())

    if dice_weight > 0:
        probs_fg = F.softmax(logits_f, dim=1)[:, num_actions - 1 : num_actions]
        inter = (probs_fg * gt_mask_f).sum()
        dice_loss = 1.0 - (2.0 * inter + 1e-6) / (probs_fg.sum() + gt_mask_f.sum() + 1e-6)
        aux_loss = aux_loss + dice_weight * dice_loss
        dice_loss_value = float(dice_loss.detach().item())

    return aux_loss, ce_loss_value, dice_loss_value

def make_optimizer(
    model: nn.Module,
    strategy: int,
    head_lr: float,
    encoder_lr: float,
    weight_decay: float,
    rl_lr: float | None = None,
):
    raw = _unwrap_compiled(model)
    encoder_params = []
    decoder_params = []
    rl_params = []

    for name, param in raw.named_parameters():
        if not param.requires_grad:
            continue
        if (
            name.startswith("extractor.encoder.")
            or name.startswith("encoder.")
            or name.startswith("smp_model.encoder.")
            or name.startswith("smp_encoder.")
        ):
            encoder_params.append(param)
        elif "decoder" in name or "segmentation_head" in name:
            decoder_params.append(param)
        else:
            rl_params.append(param)

    decoder_lr = float(_job_param("decoder_lr", head_lr))
    rl_group_lr = float(_job_param("rl_lr", rl_lr if rl_lr is not None else head_lr))

    param_groups: list[dict[str, Any]] = []

    if encoder_params:
        param_groups.append({"params": encoder_params, "lr": encoder_lr})

    if decoder_params:
        param_groups.append({"params": decoder_params, "lr": decoder_lr})

    if rl_params:
        param_groups.append({"params": rl_params, "lr": rl_group_lr})

    try:
        optimizer = AdamW(param_groups, weight_decay=weight_decay, fused=DEVICE.type == "cuda")
    except Exception:
        optimizer = AdamW(param_groups, weight_decay=weight_decay)

    return optimizer

def infer_segmentation_mask(
    model: nn.Module,
    image: torch.Tensor,
    tmax: int,
    *,
    strategy: int,
    use_amp: bool,
    amp_dtype: torch.dtype,
    use_channels_last: bool,
) -> torch.Tensor:
    model.eval()
    if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
        image = image.contiguous(memory_format=torch.channels_last)

    if strategy == 2:
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            logits = model(image)
        return threshold_binary_mask(torch.sigmoid(logits)).float()

    refinement_runtime = _uses_refinement_runtime(model, strategy=strategy)
    if strategy in (3, 4) and refinement_runtime:
        keep_index = NUM_ACTIONS // 2
        min_refinement_steps = int(_job_param("min_refinement_steps", 2))
        early_stop_keep_ratio = float(_job_param("early_stop_keep_ratio", 0.985))
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            refinement_context = model.prepare_refinement_context(image)
        seg = refinement_context["decoder_prob"].float()
        for step_idx in range(tmax):
            with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                state_t, _ = model.forward_refinement_state(
                    refinement_context["base_features"],
                    seg,
                    refinement_context["decoder_prob"],
                )
                policy_logits, _ = model.forward_from_state(state_t)
                actions, _, _ = sample_actions(policy_logits, stochastic=False)
            seg = apply_actions(
                seg,
                actions,
                soft_update_step=float(_job_param("refine_delta_small", 0.10)),
                num_actions=policy_logits.shape[1],
            ).to(dtype=seg.dtype)
            if step_idx + 1 >= min_refinement_steps:
                keep_ratio = float((actions == keep_index).float().mean().item())
                if keep_ratio >= early_stop_keep_ratio:
                    break
        return threshold_binary_mask(seg.float()).float()

    if strategy in (3, 4):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            seg = threshold_binary_mask(torch.sigmoid(model.forward_decoder(image))).float()
    else:
        seg = torch.ones(image.shape[0], 1, image.shape[2], image.shape[3], device=image.device, dtype=image.dtype)

    for _ in range(tmax):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            x_t = image * seg
            policy_logits = model.forward_policy_only(x_t)
            actions, _, _ = sample_actions(policy_logits, stochastic=False)
        seg = apply_actions(seg, actions, num_actions=policy_logits.shape[1]).to(dtype=seg.dtype)
    return seg.float()

def train_step(
    model: nn.Module,
    batch: dict[str, Any],
    optimizer: torch.optim.Optimizer,
    *,
    gamma: float,
    tmax: int,
    critic_loss_weight: float,
    log_alpha: torch.Tensor,
    alpha_optimizer: torch.optim.Optimizer,
    target_entropy: float,
    ce_weight: float,
    dice_weight: float,
    grad_clip_norm: float,
    scaler: Any | None,
    use_amp: bool,
    amp_dtype: torch.dtype,
    stepwise_backward: bool,
    use_channels_last: bool,
    initial_mask: torch.Tensor | None = None,
    decoder_loss_extra: torch.Tensor | None = None,
) -> dict[str, Any]:
    model.train()
    image = batch["image"]
    gt_mask = batch["mask"]

    if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
        image = image.contiguous(memory_format=torch.channels_last)
    if use_amp and image.device.type == "cuda":
        image = image.to(dtype=amp_dtype)
        gt_mask = gt_mask.to(dtype=amp_dtype)
    else:
        image = image.float()
        gt_mask = gt_mask.float()

    if initial_mask is not None:
        seg = initial_mask.to(device=image.device, dtype=image.dtype)
    else:
        seg = torch.ones(image.shape[0], 1, image.shape[2], image.shape[3], device=image.device, dtype=image.dtype)
    alpha = log_alpha.exp()

    total_actor = 0.0
    total_critic = 0.0
    total_loss = 0.0
    total_reward = 0.0
    total_entropy = 0.0
    total_ce_loss = 0.0
    total_dice_loss = 0.0
    accum_tensor = None
    alpha_loss_accum = torch.tensor(0.0, device=image.device, dtype=torch.float32)
    aux_fused = False

    optimizer.zero_grad(set_to_none=True)

    for _ in range(tmax):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            x_t = image * seg
            state_t, _ = model.forward_state(x_t)
            policy_logits, value_t = model.forward_from_state(state_t)
            actions, log_prob, entropy = sample_actions(policy_logits, stochastic=True)
            log_prob = log_prob.clamp(min=-10.0)
            seg_next = apply_actions(seg, actions, num_actions=policy_logits.shape[1])
            reward = (seg - gt_mask).pow(2) - (seg_next - gt_mask).pow(2)

            with torch.no_grad():
                with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                    x_next = image * seg_next
                    state_next, _ = model.forward_state(x_next)
                    value_next = model.value_from_state(state_next).detach()

            neighborhood_next = model.neighborhood_value(value_next)
            target = reward + gamma * neighborhood_next
            advantage = target - value_t
            critic_loss = F.smooth_l1_loss(value_t, target)
            actor_loss = -(log_prob * advantage.detach()).mean()
            actor_loss = actor_loss - alpha.detach() * entropy
            step_loss = (actor_loss + critic_loss_weight * critic_loss) / float(tmax)
            alpha_loss_accum = alpha_loss_accum + (log_alpha * (entropy.detach() - target_entropy)) / float(tmax)

            if not aux_fused and initial_mask is None and (ce_weight > 0 or dice_weight > 0):
                aux_loss, ce_loss_value, dice_loss_value = compute_strategy1_aux_segmentation_loss(
                    policy_logits,
                    gt_mask,
                    ce_weight=ce_weight,
                    dice_weight=dice_weight,
                    seg_mask=seg,
                )
                if ce_weight > 0:
                    total_ce_loss = ce_loss_value
                if dice_weight > 0:
                    total_dice_loss = dice_loss_value
                step_loss = step_loss + aux_loss
                aux_fused = True

        total_actor += float(actor_loss.detach().item())
        total_critic += float(critic_loss.detach().item())
        total_loss += float(step_loss.detach().item())
        total_reward += float(reward.detach().mean().item())
        total_entropy += float(entropy.detach().item())

        if stepwise_backward:
            if scaler is not None:
                scaler.scale(step_loss).backward()
            else:
                step_loss.backward()
        else:
            accum_tensor = step_loss if accum_tensor is None else accum_tensor + step_loss

        seg = seg_next.detach()

    if not stepwise_backward and accum_tensor is not None:
        if scaler is not None:
            scaler.scale(accum_tensor).backward()
        else:
            accum_tensor.backward()

    if not aux_fused and (ce_weight > 0 or dice_weight > 0):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            policy_aux, _, _ = model(image)
            logits_f = policy_aux.float()
            aux_loss = torch.zeros(1, device=image.device, dtype=torch.float32)
            if ce_weight > 0:
                num_actions = int(logits_f.shape[1])
                if num_actions >= 3:
                    init_seg = initial_mask if initial_mask is not None else torch.ones_like(gt_mask)
                    seg_bin = threshold_binary_long(init_seg).squeeze(1)
                    gt_bin = gt_mask[:, 0].long()
                    aux_target = torch.where(
                        gt_bin == 0,
                        torch.zeros_like(gt_bin),
                        torch.where(seg_bin == 1, torch.ones_like(gt_bin), torch.full_like(gt_bin, 2)),
                    )
                else:
                    aux_target = gt_mask[:, 0].long() * (num_actions - 1)
                ce_loss = F.cross_entropy(logits_f, aux_target)
                aux_loss = aux_loss + ce_weight * ce_loss
                total_ce_loss = float(ce_loss.detach().item())
            if dice_weight > 0:
                probs_fg = F.softmax(logits_f, dim=1)[:, num_actions - 1 : num_actions]
                gt_f = gt_mask.float()
                inter = (probs_fg * gt_f).sum()
                dice_loss = 1.0 - (2.0 * inter + 1e-6) / (probs_fg.sum() + gt_f.sum() + 1e-6)
                aux_loss = aux_loss + dice_weight * dice_loss
                total_dice_loss = float(dice_loss.detach().item())
        if decoder_loss_extra is not None:
            aux_loss = aux_loss + decoder_loss_extra
        if scaler is not None:
            scaler.scale(aux_loss).backward()
        else:
            aux_loss.backward()
        total_loss += float(aux_loss.detach().item())
    elif decoder_loss_extra is not None:
        if scaler is not None:
            scaler.scale(decoder_loss_extra).backward()
        else:
            decoder_loss_extra.backward()
        total_loss += float(decoder_loss_extra.detach().item())

    if scaler is not None:
        scaler.unscale_(optimizer)
    total_grad_norm = float(torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip_norm).item()) if grad_clip_norm > 0 else 0.0

    if scaler is not None:
        scaler.step(optimizer)
        scaler.update()
    else:
        optimizer.step()

    alpha_optimizer.zero_grad(set_to_none=True)
    alpha_loss_accum.backward()
    alpha_optimizer.step()
    with torch.no_grad():
        log_alpha.clamp_(min=-5.0, max=5.0)

    return {
        "loss": total_loss,
        "actor_loss": total_actor / tmax,
        "critic_loss": total_critic / tmax,
        "mean_reward": total_reward / tmax,
        "entropy": total_entropy / tmax,
        "ce_loss": total_ce_loss,
        "dice_loss": total_dice_loss,
        "grad_norm": total_grad_norm,
        "grad_clip_used": grad_clip_norm,
        "final_mask": seg.detach(),
    }

def train_step_strategy3(
    model: nn.Module,
    batch: dict[str, Any],
    optimizer: torch.optim.Optimizer,
    *,
    gamma: float,
    tmax: int,
    critic_loss_weight: float,
    log_alpha: torch.Tensor,
    alpha_optimizer: torch.optim.Optimizer,
    target_entropy: float,
    ce_weight: float,
    dice_weight: float,
    grad_clip_norm: float,
    scaler: Any | None,
    use_amp: bool,
    amp_dtype: torch.dtype,
    stepwise_backward: bool,
    use_channels_last: bool,
) -> dict[str, Any]:
    if not _uses_refinement_runtime(model, strategy=3):
        image = batch["image"]
        gt_mask = batch["mask"]
        if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
            image = image.contiguous(memory_format=torch.channels_last)
        if use_amp and image.device.type == "cuda":
            image = image.to(dtype=amp_dtype)
            gt_mask = gt_mask.to(dtype=amp_dtype)
        else:
            image = image.float()
            gt_mask = gt_mask.float()

        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            decoder_logits = model.forward_decoder(image)
        init_mask = threshold_binary_mask(torch.sigmoid(decoder_logits)).float()

        decoder_loss = torch.zeros(1, device=image.device, dtype=torch.float32)
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            dl_f = decoder_logits.float()
            gt_f = gt_mask.float()
            if ce_weight > 0:
                decoder_loss = decoder_loss + ce_weight * F.binary_cross_entropy_with_logits(dl_f, gt_f)
            if dice_weight > 0:
                dm_prob = torch.sigmoid(dl_f)
                inter = (dm_prob * gt_f).sum()
                decoder_loss = decoder_loss + dice_weight * (
                    1.0 - (2.0 * inter + 1e-6) / (dm_prob.sum() + gt_f.sum() + 1e-6)
                )

        return train_step(
            model,
            batch,
            optimizer,
            gamma=gamma,
            tmax=tmax,
            critic_loss_weight=critic_loss_weight,
            log_alpha=log_alpha,
            alpha_optimizer=alpha_optimizer,
            target_entropy=target_entropy,
            ce_weight=ce_weight,
            dice_weight=dice_weight,
            grad_clip_norm=grad_clip_norm,
            scaler=scaler,
            use_amp=use_amp,
            amp_dtype=amp_dtype,
            stepwise_backward=stepwise_backward,
            use_channels_last=use_channels_last,
            initial_mask=init_mask,
            decoder_loss_extra=decoder_loss,
        )

    model.train()
    image = batch["image"]
    gt_mask = batch["mask"]
    if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
        image = image.contiguous(memory_format=torch.channels_last)
    if use_amp and image.device.type == "cuda":
        image = image.to(dtype=amp_dtype)
        gt_mask = gt_mask.to(dtype=amp_dtype)
    else:
        image = image.float()
        gt_mask = gt_mask.float()

    del stepwise_backward

    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
        refinement_context = model.prepare_refinement_context(image)
    decoder_logits = refinement_context["decoder_logits"]
    decoder_prob = refinement_context["decoder_prob"]
    base_features = refinement_context["base_features"]
    seg = decoder_prob.detach().to(device=image.device, dtype=image.dtype)

    decoder_loss = torch.zeros((), device=image.device, dtype=torch.float32)
    decoder_ce_loss_value = 0.0
    decoder_dice_loss_value = 0.0
    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
        dl_f = decoder_logits.float()
        gt_f = gt_mask.float()
        if ce_weight > 0:
            decoder_ce = F.binary_cross_entropy_with_logits(dl_f, gt_f)
            decoder_loss = decoder_loss + ce_weight * decoder_ce
            decoder_ce_loss_value = float(decoder_ce.detach().item())
        if dice_weight > 0:
            dm_prob = decoder_prob.float()
            inter = (dm_prob * gt_f).sum()
            decoder_dice = 1.0 - (2.0 * inter + 1e-6) / (dm_prob.sum() + gt_f.sum() + 1e-6)
            decoder_loss = decoder_loss + dice_weight * decoder_dice
            decoder_dice_loss_value = float(decoder_dice.detach().item())

    optimizer.zero_grad(set_to_none=True)

    alpha = log_alpha.exp()
    gae_lambda = float(_job_param("gae_lambda", 0.90))
    min_refinement_steps = int(_job_param("min_refinement_steps", 2))
    early_stop_keep_ratio = float(_job_param("early_stop_keep_ratio", 0.985))
    keep_index = NUM_ACTIONS // 2

    transitions: list[dict[str, Any]] = []
    alpha_loss_accum = torch.zeros((), device=image.device, dtype=torch.float32)
    decoder_prior = decoder_prob.detach()
    detached_base_features = base_features.detach()
    detached_decoder_prob = decoder_prob.detach()

    for step_idx in range(tmax):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            state_t, _ = model.forward_refinement_state(
                base_features,
                seg,
                decoder_prob,
            )
            policy_logits, value_t = model.forward_from_state(state_t)
            actions, log_prob, entropy = sample_actions(policy_logits, stochastic=True)
            log_prob = log_prob.clamp(min=-10.0)
            seg_next = apply_actions(
                seg,
                actions,
                soft_update_step=float(_job_param("refine_delta_small", 0.10)),
                num_actions=policy_logits.shape[1],
            )
            reward = compute_refinement_reward(seg, seg_next, gt_mask.float(), decoder_prior=decoder_prior)
            aux_loss, ce_loss_value, dice_loss_value = compute_strategy1_aux_segmentation_loss(
                policy_logits,
                gt_mask,
                ce_weight=ce_weight,
                dice_weight=dice_weight,
                seg_mask=seg,
            )

            with torch.no_grad():
                state_next, _ = model.forward_refinement_state(
                    detached_base_features,
                    seg_next,
                    detached_decoder_prob,
                )
                value_next = model.value_from_state(state_next).detach()

        transitions.append(
            {
                "value": value_t,
                "next_value": value_next,
                "reward": reward,
                "log_prob": log_prob,
                "entropy": entropy,
                "aux_loss": aux_loss,
                "ce_loss": ce_loss_value,
                "dice_loss": dice_loss_value,
            }
        )
        alpha_loss_accum = alpha_loss_accum + (log_alpha * (entropy.detach() - target_entropy))
        seg = seg_next.detach()

        if step_idx + 1 >= min_refinement_steps:
            keep_ratio = float((actions == keep_index).float().mean().item())
            if keep_ratio >= early_stop_keep_ratio:
                break

    effective_steps = max(len(transitions), 1)
    advantages_rev: list[torch.Tensor] = []
    gae = torch.zeros_like(transitions[-1]["value"])
    for transition in reversed(transitions):
        td_error = transition["reward"] + gamma * model.neighborhood_value(transition["next_value"]) - transition["value"]
        gae = td_error + gamma * gae_lambda * gae
        advantages_rev.append(gae)
    advantages = list(reversed(advantages_rev))

    total_loss_tensor = decoder_loss
    total_actor = 0.0
    total_critic = 0.0
    total_reward = 0.0
    total_entropy = 0.0
    total_ce_loss = decoder_ce_loss_value
    total_dice_loss = decoder_dice_loss_value

    for transition, advantage in zip(transitions, advantages):
        detached_adv = advantage.detach()
        norm_adv = (detached_adv - detached_adv.mean()) / (detached_adv.std(unbiased=False) + 1e-6)
        critic_target = (transition["value"] + detached_adv).detach()
        critic_loss = F.smooth_l1_loss(transition["value"], critic_target)
        actor_loss = -(transition["log_prob"] * norm_adv).mean()
        actor_loss = actor_loss - alpha.detach() * transition["entropy"]
        step_loss = (actor_loss + critic_loss_weight * critic_loss + transition["aux_loss"]) / float(effective_steps)
        total_loss_tensor = total_loss_tensor + step_loss

        total_actor += float(actor_loss.detach().item())
        total_critic += float(critic_loss.detach().item())
        total_reward += float(transition["reward"].detach().mean().item())
        total_entropy += float(transition["entropy"].detach().item())
        total_ce_loss += float(transition["ce_loss"]) / float(effective_steps)
        total_dice_loss += float(transition["dice_loss"]) / float(effective_steps)

    if scaler is not None:
        scaler.scale(total_loss_tensor).backward()
        scaler.unscale_(optimizer)
    else:
        total_loss_tensor.backward()

    total_grad_norm = (
        float(torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip_norm).item()) if grad_clip_norm > 0 else 0.0
    )

    if scaler is not None:
        scaler.step(optimizer)
        scaler.update()
    else:
        optimizer.step()

    alpha_optimizer.zero_grad(set_to_none=True)
    (alpha_loss_accum / float(effective_steps)).backward()
    alpha_optimizer.step()
    with torch.no_grad():
        log_alpha.clamp_(min=-5.0, max=5.0)

    return {
        "loss": float(total_loss_tensor.detach().item()),
        "actor_loss": total_actor / float(effective_steps),
        "critic_loss": total_critic / float(effective_steps),
        "mean_reward": total_reward / float(effective_steps),
        "entropy": total_entropy / float(effective_steps),
        "ce_loss": total_ce_loss,
        "dice_loss": total_dice_loss,
        "grad_norm": total_grad_norm,
        "grad_clip_used": grad_clip_norm,
        "final_mask": threshold_binary_mask(seg.detach().float()).float(),
    }

def train_step_supervised(
    model: nn.Module,
    batch: dict[str, Any],
    optimizer: torch.optim.Optimizer,
    *,
    scaler: Any | None,
    use_amp: bool,
    amp_dtype: torch.dtype,
    use_channels_last: bool,
    grad_clip_norm: float,
    ce_weight: float = 0.5,
    dice_weight: float = 0.5,
) -> dict[str, Any]:
    model.train()
    image = batch["image"]
    gt_mask = batch["mask"]

    if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
        image = image.contiguous(memory_format=torch.channels_last)
    if use_amp and image.device.type == "cuda":
        image = image.to(dtype=amp_dtype)
        gt_mask = gt_mask.to(dtype=amp_dtype)
    else:
        image = image.float()
        gt_mask = gt_mask.float()

    optimizer.zero_grad(set_to_none=True)
    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
        logits = model(image)
        logits_f = logits.float()
        gt_f = gt_mask.float()
        loss = torch.zeros(1, device=image.device, dtype=torch.float32)
        ce_loss_val = 0.0
        dice_loss_val = 0.0
        if ce_weight > 0:
            bce = F.binary_cross_entropy_with_logits(logits_f, gt_f, reduction="mean")
            loss = loss + ce_weight * bce
            ce_loss_val = float(bce.detach().item())
        if dice_weight > 0:
            pred_f = torch.sigmoid(logits_f)
            inter = (pred_f * gt_f).sum()
            dice_l = 1.0 - (2.0 * inter + 1e-6) / (pred_f.sum() + gt_f.sum() + 1e-6)
            loss = loss + dice_weight * dice_l
            dice_loss_val = float(dice_l.detach().item())

    if scaler is not None:
        scaler.scale(loss).backward()
        scaler.unscale_(optimizer)
    else:
        loss.backward()

    total_grad_norm = float(torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip_norm).item()) if grad_clip_norm > 0 else 0.0
    if scaler is not None:
        scaler.step(optimizer)
        scaler.update()
    else:
        optimizer.step()

    final_mask = threshold_binary_mask(torch.sigmoid(logits_f)).float().detach()
    return {
        "loss": float(loss.detach().item()),
        "actor_loss": 0.0,
        "critic_loss": 0.0,
        "mean_reward": 0.0,
        "entropy": 0.0,
        "ce_loss": ce_loss_val,
        "dice_loss": dice_loss_val,
        "grad_norm": total_grad_norm,
        "grad_clip_used": grad_clip_norm,
        "final_mask": final_mask,
    }

@torch.inference_mode()
def validate(
    model: nn.Module,
    loader: DataLoader,
    *,
    strategy: int,
    tmax: int,
    use_amp: bool,
    amp_dtype: torch.dtype,
    use_channels_last: bool,
    gamma: float,
    critic_loss_weight: float,
    ce_weight: float,
    dice_weight: float,
) -> dict[str, float]:
    model.eval()
    losses: list[float] = []
    dice_scores: list[float] = []
    iou_scores: list[float] = []
    entropies: list[float] = []
    rewards: list[float] = []
    actor_losses: list[float] = []
    critic_losses: list[float] = []
    ce_losses: list[float] = []
    dice_losses: list[float] = []
    prefetcher = CUDAPrefetcher(loader, DEVICE)
    try:
        for batch in tqdm(prefetcher, total=len(loader), desc="Validating", leave=False):
            image = batch["image"]
            gt_mask = batch["mask"].float()
            if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
                image = image.contiguous(memory_format=torch.channels_last)
            if use_amp and DEVICE.type == "cuda":
                image = image.to(dtype=amp_dtype)
                gt_mask = gt_mask.to(dtype=amp_dtype)

            if strategy == 2:
                with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                    logits = model(image)
                pred_prob = torch.sigmoid(logits).float()
                pred = threshold_binary_mask(pred_prob).float()
                bce = F.binary_cross_entropy_with_logits(logits.float(), gt_mask.float(), reduction="mean") if ce_weight > 0 else torch.tensor(0.0, device=image.device)
                inter_prob = (pred_prob * gt_mask.float()).sum()
                dice_loss = 1.0 - (2.0 * inter_prob + 1e-6) / (pred_prob.sum() + gt_mask.sum() + 1e-6) if dice_weight > 0 else torch.tensor(0.0, device=image.device)
                losses.append(float((ce_weight * bce + dice_weight * dice_loss).item()))
                ce_losses.append(float(bce.item()) if ce_weight > 0 else 0.0)
                dice_losses.append(float(dice_loss.item()) if dice_weight > 0 else 0.0)
            else:
                refinement_runtime = _uses_refinement_runtime(model, strategy=strategy)
                if strategy in (3, 4) and refinement_runtime:
                    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                        refinement_context = model.prepare_refinement_context(image)
                    decoder_logits = refinement_context["decoder_logits"]
                    decoder_prob = refinement_context["decoder_prob"].float()
                    seg = decoder_prob.float()
                    decoder_loss = 0.0
                    batch_ce = 0.0
                    batch_dice = 0.0
                    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                        dl_f = decoder_logits.float()
                        gt_f = gt_mask.float()
                        if ce_weight > 0:
                            batch_ce = float(F.binary_cross_entropy_with_logits(dl_f, gt_f).item())
                            decoder_loss += ce_weight * batch_ce
                        if dice_weight > 0:
                            dm_prob = decoder_prob.float()
                            inter = (dm_prob * gt_f).sum()
                            batch_dice = float(
                                (
                                    1.0 - (2.0 * inter + 1e-6) / (dm_prob.sum() + gt_f.sum() + 1e-6)
                                ).item()
                            )
                            decoder_loss += dice_weight * batch_dice
                else:
                    fg_count = gt_mask.sum().clamp(min=1.0)
                    bg_count = (gt_mask == 0).sum().clamp(min=1.0)
                    pos_weight = bg_count / fg_count
                    _weight_map = torch.where(gt_mask == 1, pos_weight, torch.ones_like(gt_mask))
                    if strategy in (3, 4):
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            decoder_logits = model.forward_decoder(image)
                        seg = threshold_binary_mask(torch.sigmoid(decoder_logits)).float()
                        decoder_loss = 0.0
                        batch_ce = 0.0
                        batch_dice = 0.0
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            dl_f = decoder_logits.float()
                            gt_f = gt_mask.float()
                            if ce_weight > 0:
                                batch_ce = float(F.binary_cross_entropy_with_logits(dl_f, gt_f).item())
                                decoder_loss += ce_weight * batch_ce
                            if dice_weight > 0:
                                dm_prob = torch.sigmoid(dl_f)
                                inter = (dm_prob * gt_f).sum()
                                batch_dice = float(
                                    (
                                        1.0 - (2.0 * inter + 1e-6) / (dm_prob.sum() + gt_f.sum() + 1e-6)
                                    ).item()
                                )
                                decoder_loss += dice_weight * batch_dice
                    else:
                        seg = torch.ones(
                            image.shape[0],
                            1,
                            image.shape[2],
                            image.shape[3],
                            device=image.device,
                            dtype=image.dtype,
                        )
                        decoder_loss = 0.0
                        batch_ce = 0.0
                        batch_dice = 0.0

                batch_actor = 0.0
                batch_critic = 0.0
                batch_reward = 0.0
                batch_entropy = 0.0
                batch_loss = decoder_loss

                if strategy in (3, 4) and refinement_runtime:
                    decoder_prior = decoder_prob.detach()
                    detached_base_features = refinement_context["base_features"].detach()
                    detached_decoder_prob = decoder_prob.detach()

                    for _ in range(tmax):
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            state_t, _ = model.forward_refinement_state(
                                refinement_context["base_features"],
                                seg,
                                decoder_prob,
                            )
                            policy_logits, value_t = model.forward_from_state(state_t)
                            actions, log_prob, entropy = sample_actions(policy_logits, stochastic=False)
                            log_prob = log_prob.clamp(min=-10.0)
                            seg_next = apply_actions(seg, actions, soft_update_step=DEFAULT_MASK_UPDATE_STEP, num_actions=policy_logits.shape[1])
                            reward = compute_refinement_reward(
                                seg,
                                seg_next,
                                gt_mask.float(),
                                decoder_prior=decoder_prior,
                            )
                            state_next, _ = model.forward_refinement_state(
                                detached_base_features,
                                seg_next,
                                detached_decoder_prob,
                            )
                            value_next = model.value_from_state(state_next).detach()
                            target = reward + gamma * model.neighborhood_value(value_next)
                            advantage = target - value_t
                            norm_advantage = advantage.detach()
                            norm_advantage = (norm_advantage - norm_advantage.mean()) / (
                                norm_advantage.std(unbiased=False) + 1e-6
                            )
                            actor_loss = -(log_prob * norm_advantage).mean()
                            critic_loss = F.smooth_l1_loss(value_t, target)
                            aux_loss, aux_ce, aux_dice = compute_strategy1_aux_segmentation_loss(
                                policy_logits,
                                gt_mask,
                                ce_weight=ce_weight,
                                dice_weight=dice_weight,
                                seg_mask=seg,
                            )

                        batch_actor += float(actor_loss.item())
                        batch_critic += float(critic_loss.item())
                        batch_reward += float(reward.mean().item())
                        batch_entropy += float(entropy.item())
                        batch_ce += aux_ce / float(tmax)
                        batch_dice += aux_dice / float(tmax)
                        batch_loss += float(
                            ((actor_loss + critic_loss_weight * critic_loss + aux_loss) / float(tmax)).item()
                        )
                        seg = seg_next
                    pred = threshold_binary_mask(seg.float()).float()
                else:
                    for _ in range(tmax):
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            x_t = image * seg
                            state_t, _ = model.forward_state(x_t)
                            policy_logits, value_t = model.forward_from_state(state_t)
                            actions, log_prob, entropy = sample_actions(policy_logits, stochastic=False)
                            log_prob = log_prob.clamp(min=-10.0)
                            seg_next = apply_actions(seg, actions, num_actions=policy_logits.shape[1])
                            reward = ((seg - gt_mask).pow(2) - (seg_next - gt_mask).pow(2))
                            x_next = image * seg_next
                            state_next, _ = model.forward_state(x_next)
                            value_next = model.value_from_state(state_next).detach()
                            target = reward + gamma * model.neighborhood_value(value_next)
                            advantage = target - value_t
                            actor_loss = -(log_prob * advantage.detach()).mean()
                            critic_loss = F.smooth_l1_loss(value_t, target)

                        batch_actor += float(actor_loss.item())
                        batch_critic += float(critic_loss.item())
                        batch_reward += float(reward.mean().item())
                        batch_entropy += float(entropy.item())
                        batch_loss += float(((actor_loss + critic_loss_weight * critic_loss) / float(tmax)).item())
                        seg = seg_next

                    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                        policy_aux, _, _ = model(image)
                        aux_loss, aux_ce, aux_dice = compute_strategy1_aux_segmentation_loss(
                            policy_aux,
                            gt_mask,
                            ce_weight=ce_weight,
                            dice_weight=dice_weight,
                            seg_mask=seg,
                        )
                    batch_loss += float(aux_loss.item())
                    batch_ce = aux_ce
                    batch_dice = aux_dice
                    pred = infer_segmentation_mask(
                        model,
                        image,
                        tmax,
                        strategy=strategy,
                        use_amp=use_amp,
                        amp_dtype=amp_dtype,
                        use_channels_last=use_channels_last,
                    ).float()

                ce_losses.append(batch_ce)
                dice_losses.append(batch_dice)
                actor_losses.append(batch_actor / tmax)
                critic_losses.append(batch_critic / tmax)
                rewards.append(batch_reward / tmax)
                entropies.append(batch_entropy / tmax)
                losses.append(batch_loss)

            inter = (pred * gt_mask.float()).sum(dim=(1, 2, 3))
            pred_sum = pred.sum(dim=(1, 2, 3))
            gt_sum = gt_mask.float().sum(dim=(1, 2, 3))
            dice = (2.0 * inter + _EPS) / (pred_sum + gt_sum + _EPS)
            iou = (inter + _EPS) / (pred_sum + gt_sum - inter + _EPS)
            dice_scores.extend(dice.cpu().tolist())
            iou_scores.extend(iou.cpu().tolist())
    finally:
        prefetcher.close()
        del prefetcher

    return {
        "val_loss": float(np.mean(losses)) if losses else 0.0,
        "val_dice": float(np.mean(dice_scores)) if dice_scores else 0.0,
        "val_iou": float(np.mean(iou_scores)) if iou_scores else 0.0,
        "val_actor_loss": float(np.mean(actor_losses)) if actor_losses else 0.0,
        "val_critic_loss": float(np.mean(critic_losses)) if critic_losses else 0.0,
        "val_ce_loss": float(np.mean(ce_losses)) if ce_losses else 0.0,
        "val_dice_loss": float(np.mean(dice_losses)) if dice_losses else 0.0,
        "val_reward": float(np.mean(rewards)) if rewards else 0.0,
        "val_entropy": float(np.mean(entropies)) if entropies else 0.0,
    }

def _save_training_plots(history: list[dict[str, Any]], plots_dir: Path) -> None:
    if len(history) < 1:
        return
    ensure_dir(plots_dir)
    epochs = [row["epoch"] for row in history]
    plot_specs = [
        ("loss.png", "Loss", [("train_loss", "Train"), ("val_loss", "Val")]),
        ("dice.png", "Dice", [("train_dice", "Train"), ("val_dice", "Val")]),
        ("iou.png", "IoU", [("train_iou", "Train"), ("val_iou", "Val")]),
        ("reward.png", "Reward", [("train_mean_reward", "Train"), ("val_reward", "Val")]),
        ("entropy.png", "Entropy", [("train_entropy", "Train"), ("val_entropy", "Val")]),
        ("ce_loss.png", "CE Loss", [("train_ce_loss", "Train")]),
        ("dice_loss.png", "Dice Loss", [("train_dice_loss", "Train")]),
        ("alpha.png", "Alpha", [("alpha", "Alpha")]),
        ("lr.png", "Learning Rate", [("lr", "Head LR"), ("encoder_lr", "Encoder LR")]),
    ]
    for file_name, title, curves in plot_specs:
        fig, ax = plt.subplots(figsize=(8, 4))
        has_data = False
        for key, label in curves:
            values = [(row["epoch"], row[key]) for row in history if key in row]
            if not values:
                continue
            xs, ys = zip(*values)
            ax.plot(xs, ys, label=label, linewidth=1.2)
            has_data = True
        if has_data:
            ax.set_title(title)
            ax.set_xlabel("Epoch")
            ax.set_ylabel(title)
            ax.grid(True, alpha=0.3)
            ax.legend()
            fig.tight_layout()
            fig.savefig(plots_dir / file_name, dpi=110)
        plt.close(fig)

RESUME_IDENTITY_KEYS = (
    "strategy",
    "dataset_percent",
    "dataset_name",
    "dataset_split_policy",
    "split_type",
    "train_subset_key",
    "backbone_family",
    "smp_encoder_name",
    "smp_encoder_weights",
    "smp_encoder_depth",
    "smp_encoder_proj_dim",
    "smp_decoder_type",
    "vgg_feature_scales",
    "vgg_feature_dilation",
    "head_lr",
    "encoder_lr",
    "weight_decay",
    "dropout_p",
    "tmax",
    "entropy_lr",
)

def _resume_value_matches(current: Any, saved: Any) -> bool:
    if isinstance(current, (int, float)) and isinstance(saved, (int, float)) and not isinstance(current, bool):
        return math.isclose(float(current), float(saved), rel_tol=1e-9, abs_tol=1e-12)
    return current == saved

def validate_resume_checkpoint_identity(
    current_run_config: dict[str, Any],
    saved_run_config: dict[str, Any],
    *,
    checkpoint_path: Path,
) -> None:
    mismatches: list[str] = []
    for key in RESUME_IDENTITY_KEYS:
        if key not in current_run_config or key not in saved_run_config:
            mismatches.append(f"{key}: current={current_run_config.get(key)!r}, checkpoint={saved_run_config.get(key)!r}")
            continue
        if not _resume_value_matches(current_run_config[key], saved_run_config[key]):
            mismatches.append(f"{key}: current={current_run_config[key]!r}, checkpoint={saved_run_config[key]!r}")

    current_s2 = current_run_config.get("strategy2_checkpoint_path")
    saved_s2 = saved_run_config.get("strategy2_checkpoint_path")
    if current_s2 or saved_s2:
        if str(current_s2 or "") != str(saved_s2 or ""):
            mismatches.append(f"strategy2_checkpoint_path: current={current_s2!r}, checkpoint={saved_s2!r}")

    if mismatches:
        raise ValueError(
            f"Resume checkpoint identity mismatch for {checkpoint_path}:\n" + "\n".join(f"  - {line}" for line in mismatches)
        )

def load_history_for_resume(history_path: Path, checkpoint_payload: dict[str, Any]) -> list[dict[str, Any]]:
    checkpoint_epoch = int(checkpoint_payload.get("epoch", 0))
    epoch_metrics = checkpoint_payload.get("epoch_metrics")
    history: list[dict[str, Any]] = []
    if history_path.exists():
        payload = load_json(history_path)
        if not isinstance(payload, list):
            raise RuntimeError(f"Expected list history at {history_path}, found {type(payload).__name__}.")
        history = [dict(row) for row in payload if isinstance(row, dict)]
        history = [row for row in history if int(row.get("epoch", 0)) <= checkpoint_epoch]
    if not history and isinstance(epoch_metrics, dict):
        history = [dict(epoch_metrics)]
    elif history and isinstance(epoch_metrics, dict):
        if int(history[-1].get("epoch", 0)) < checkpoint_epoch:
            history.append(dict(epoch_metrics))
    return history

def train_model(
    *,
    run_type: str,
    model_config: RuntimeModelConfig,
    run_config: dict[str, Any],
    model: nn.Module,
    description: str,
    strategy: int,
    run_dir: Path,
    bundle: DataBundle,
    max_epochs: int,
    head_lr: float,
    encoder_lr: float,
    weight_decay: float,
    tmax: int,
    entropy_lr: float,
    entropy_alpha_init: float,
    entropy_target_ratio: float,
    critic_loss_weight: float,
    ce_weight: float,
    dice_weight: float,
    dropout_p: float,
    resume_checkpoint_path: Path | None = None,
    trial: optuna.trial.Trial | None = None,
) -> tuple[dict[str, Any], list[dict[str, Any]]]:
    amp_dtype = resolve_amp_dtype(AMP_DTYPE)
    use_amp = amp_autocast_enabled(DEVICE)
    use_channels_last = _use_channels_last_for_run(model_config)
    if resume_checkpoint_path is not None and strategy in (3, 4):
        _configure_model_from_checkpoint_path(model, resume_checkpoint_path)
    print_model_parameter_summary(
        model=model,
        description=description,
        strategy=strategy,
        model_config=model_config,
        dropout_p=dropout_p,
        amp_dtype=amp_dtype,
        compiled=hasattr(model, "_orig_mod"),
    )

    optimizer = make_optimizer(model, strategy, head_lr=head_lr * 0.1, encoder_lr=encoder_lr, weight_decay=weight_decay, rl_lr=head_lr)
    scheduler = CosineAnnealingLR(
      optimizer,
      T_max=200,      # full training epochs
      eta_min=1e-6,   # floor
    )
    scaler = make_grad_scaler(enabled=use_amp, amp_dtype=amp_dtype, device=DEVICE)

    target_entropy = entropy_target_ratio * math.log(max(_model_policy_action_count(model) or NUM_ACTIONS, 2))
    log_alpha = torch.tensor(math.log(max(entropy_alpha_init, 1e-8)), dtype=torch.float32, device=DEVICE, requires_grad=True)
    alpha_optimizer = Adam([log_alpha], lr=entropy_lr)

    save_artifacts = run_type in {"final", "trial"}
    ckpt_dir = ensure_dir(run_dir / "checkpoints") if save_artifacts else None
    plots_dir = ensure_dir(run_dir / "plots") if save_artifacts else None
    history_path = checkpoint_history_path(run_dir, run_type)
    history: list[dict[str, Any]] = []
    best_val_iou = -1.0
    patience_counter = 0
    elapsed_before_resume = 0.0
    start_epoch = 1
    resume_source: dict[str, Any] | None = None
    if resume_checkpoint_path is not None:
        checkpoint_payload = load_checkpoint(
            resume_checkpoint_path,
            model=model,
            optimizer=optimizer,
            scheduler=scheduler,
            scaler=scaler,
            device=DEVICE,
            log_alpha=log_alpha if strategy != 2 else None,
            alpha_optimizer=alpha_optimizer if strategy != 2 else None,
            expected_run_type=run_type,
            require_run_metadata=True,
        )
        validate_resume_checkpoint_identity(
            run_config,
            checkpoint_run_config_payload(checkpoint_payload),
            checkpoint_path=resume_checkpoint_path,
        )
        start_epoch = int(checkpoint_payload["epoch"]) + 1
        best_val_iou = float(checkpoint_payload["best_metric_value"])
        patience_counter = int(checkpoint_payload.get("patience_counter", 0))
        elapsed_before_resume = float(checkpoint_payload.get("elapsed_seconds", 0.0))
        history = load_history_for_resume(history_path, checkpoint_payload)
        resume_source = {
            "checkpoint_path": str(Path(resume_checkpoint_path).resolve()),
            "checkpoint_epoch": int(checkpoint_payload["epoch"]),
            "checkpoint_run_type": checkpoint_payload.get("run_type", run_type),
        }
        print(
            f"[Resume] {run_type} run continuing from {resume_checkpoint_path} "
            f"at epoch {start_epoch}/{max_epochs}."
        )
    start_time = time.time()
    validate_interval = max(int(VALIDATE_EVERY_N_EPOCHS), 1)
    run_label = f"Trial {trial.number:03d}" if trial is not None else "Run"

    for epoch in range(start_epoch, max_epochs + 1):
        epoch_losses: list[float] = []
        epoch_actor: list[float] = []
        epoch_critic: list[float] = []
        epoch_reward: list[float] = []
        epoch_entropy: list[float] = []
        epoch_ce: list[float] = []
        epoch_dice_loss: list[float] = []
        epoch_grad: list[float] = []
        epoch_dices: list[float] = []
        epoch_ious: list[float] = []

        prefetcher = CUDAPrefetcher(bundle.train_loader, DEVICE)
        progress = tqdm(prefetcher, total=len(bundle.train_loader), desc=f"Epoch {epoch}/{max_epochs}", leave=False)
        for batch in progress:
            if strategy == 2:
                metrics = train_step_supervised(
                    model,
                    batch,
                    optimizer,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    use_channels_last=use_channels_last,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    ce_weight=ce_weight,
                    dice_weight=dice_weight,
                )
            elif strategy in (3, 4):
                metrics = train_step_strategy3(
                    model,
                    batch,
                    optimizer,
                    gamma=DEFAULT_GAMMA,
                    tmax=tmax,
                    critic_loss_weight=critic_loss_weight,
                    log_alpha=log_alpha,
                    alpha_optimizer=alpha_optimizer,
                    target_entropy=target_entropy,
                    ce_weight=ce_weight,
                    dice_weight=dice_weight,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    stepwise_backward=STEPWISE_BACKWARD,
                    use_channels_last=use_channels_last,
                )
            else:
                metrics = train_step(
                    model,
                    batch,
                    optimizer,
                    gamma=DEFAULT_GAMMA,
                    tmax=tmax,
                    critic_loss_weight=critic_loss_weight,
                    log_alpha=log_alpha,
                    alpha_optimizer=alpha_optimizer,
                    target_entropy=target_entropy,
                    ce_weight=ce_weight,
                    dice_weight=dice_weight,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    stepwise_backward=STEPWISE_BACKWARD,
                    use_channels_last=use_channels_last,
                )

            epoch_losses.append(metrics["loss"])
            epoch_actor.append(metrics["actor_loss"])
            epoch_critic.append(metrics["critic_loss"])
            epoch_reward.append(metrics["mean_reward"])
            epoch_entropy.append(metrics["entropy"])
            epoch_ce.append(metrics["ce_loss"])
            epoch_dice_loss.append(metrics["dice_loss"])
            epoch_grad.append(metrics["grad_norm"])

            pred_mask = metrics["final_mask"]
            gt_mask = batch["mask"].float()
            inter = (pred_mask * gt_mask).sum(dim=(1, 2, 3))
            pred_sum = pred_mask.sum(dim=(1, 2, 3))
            gt_sum = gt_mask.sum(dim=(1, 2, 3))
            dice = (2.0 * inter + _EPS) / (pred_sum + gt_sum + _EPS)
            iou = (inter + _EPS) / (pred_sum + gt_sum - inter + _EPS)
            epoch_dices.extend(dice.detach().cpu().tolist())
            epoch_ious.extend(iou.detach().cpu().tolist())

            head_lr_now = float(optimizer.param_groups[-1]["lr"])
            enc_lr_now = float(optimizer.param_groups[0]["lr"]) if len(optimizer.param_groups) > 1 else head_lr_now
            progress.set_postfix(loss=f"{metrics['loss']:.4f}", iou=f"{np.mean(epoch_ious):.4f}", lr=f"{head_lr_now:.2e}")

        should_validate = epoch % validate_interval == 0 or epoch == max_epochs
        val_metrics: dict[str, float | None] = {
            "val_loss": None,
            "val_dice": None,
            "val_iou": None,
            "val_reward": None,
            "val_entropy": None,
        }
        if should_validate:
            tqdm.write(
                f"[{run_label}] Epoch {epoch}/{max_epochs}: running validation on {len(bundle.val_loader)} batches..."
            )
            validated_metrics = validate(
                model,
                bundle.val_loader,
                strategy=strategy,
                tmax=tmax,
                use_amp=use_amp,
                amp_dtype=amp_dtype,
                use_channels_last=use_channels_last,
                gamma=DEFAULT_GAMMA,
                critic_loss_weight=critic_loss_weight,
                ce_weight=ce_weight,
                dice_weight=dice_weight,
            )
            val_metrics.update(validated_metrics)
        scheduler.step() # always step up the scheduler

        row = {
            "epoch": epoch,
            "train_loss": float(np.mean(epoch_losses)) if epoch_losses else 0.0,
            "train_actor_loss": float(np.mean(epoch_actor)) if epoch_actor else 0.0,
            "train_critic_loss": float(np.mean(epoch_critic)) if epoch_critic else 0.0,
            "train_mean_reward": float(np.mean(epoch_reward)) if epoch_reward else 0.0,
            "train_entropy": float(np.mean(epoch_entropy)) if epoch_entropy else 0.0,
            "train_ce_loss": float(np.mean(epoch_ce)) if epoch_ce else 0.0,
            "train_dice_loss": float(np.mean(epoch_dice_loss)) if epoch_dice_loss else 0.0,
            "train_dice": float(np.mean(epoch_dices)) if epoch_dices else 0.0,
            "train_iou": float(np.mean(epoch_ious)) if epoch_ious else 0.0,
            "grad_norm": float(np.mean(epoch_grad)) if epoch_grad else 0.0,
            "lr": float(optimizer.param_groups[-1]["lr"]),
            "encoder_lr": float(optimizer.param_groups[0]["lr"]) if len(optimizer.param_groups) > 1 else float(optimizer.param_groups[-1]["lr"]),
            "alpha": float(log_alpha.exp().detach().item()) if strategy != 2 else 0.0,
            "validated_this_epoch": should_validate,
            **val_metrics,
        }
        history.append(row)
        if save_artifacts and plots_dir is not None:
            _save_training_plots(history, plots_dir)
            save_json(history_path, history)

        improved = False
        if should_validate:
            val_iou = float(val_metrics["val_iou"])
            improved = val_iou > best_val_iou
            if improved:
                best_val_iou = val_iou
                patience_counter = 0
                if save_artifacts and ckpt_dir is not None:
                    save_checkpoint(
                        ckpt_dir / "best.pt",
                        run_type=run_type,
                        model=model,
                        optimizer=optimizer,
                        scheduler=scheduler,
                        scaler=scaler,
                        epoch=epoch,
                        best_metric_value=best_val_iou,
                        best_metric_name=BEST_CHECKPOINT_METRIC,
                        run_config=run_config,
                        epoch_metrics=row,
                        patience_counter=patience_counter,
                        elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                        log_alpha=log_alpha if strategy != 2 else None,
                        alpha_optimizer=alpha_optimizer if strategy != 2 else None,
                        resume_source=resume_source,
                    )
            else:
                patience_counter += 1

        if save_artifacts and ckpt_dir is not None and SAVE_LATEST_EVERY_EPOCH:
            save_checkpoint(
                ckpt_dir / "latest.pt",
                run_type=run_type,
                model=model,
                optimizer=optimizer,
                scheduler=scheduler,
                scaler=scaler,
                epoch=epoch,
                best_metric_value=best_val_iou,
                best_metric_name=BEST_CHECKPOINT_METRIC,
                run_config=run_config,
                epoch_metrics=row,
                patience_counter=patience_counter,
                elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                log_alpha=log_alpha if strategy != 2 else None,
                alpha_optimizer=alpha_optimizer if strategy != 2 else None,
                resume_source=resume_source,
            )
        if save_artifacts and ckpt_dir is not None and CHECKPOINT_EVERY_N_EPOCHS > 0 and epoch % CHECKPOINT_EVERY_N_EPOCHS == 0:
            save_checkpoint(
                ckpt_dir / f"epoch_{epoch:04d}.pt",
                run_type=run_type,
                model=model,
                optimizer=optimizer,
                scheduler=scheduler,
                scaler=scaler,
                epoch=epoch,
                best_metric_value=best_val_iou,
                best_metric_name=BEST_CHECKPOINT_METRIC,
                run_config=run_config,
                epoch_metrics=row,
                patience_counter=patience_counter,
                elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                log_alpha=log_alpha if strategy != 2 else None,
                alpha_optimizer=alpha_optimizer if strategy != 2 else None,
                resume_source=resume_source,
            )

        if trial is not None and should_validate:
            val_iou = float(val_metrics["val_iou"])
            trial.report(val_iou, step=epoch)
            if USE_TRIAL_PRUNING and epoch >= TRIAL_PRUNER_WARMUP_STEPS and trial.should_prune():
                raise optuna.TrialPruned(f"Trial pruned at epoch {epoch} with val_iou={val_iou:.4f}")

        if VERBOSE_EPOCH_LOG:
            tqdm.write(f"[{run_label}] Epoch {epoch}/{max_epochs}")
            tqdm.write(json.dumps(row, indent=2))
        else:
            tqdm.write(
                f"[{run_label}] Epoch {epoch}/{max_epochs}: "
                f"{format_concise_epoch_log(row, best_val_iou=best_val_iou)}"
            )

        if should_validate and EARLY_STOPPING_PATIENCE > 0 and patience_counter >= EARLY_STOPPING_PATIENCE:
            print(f"Early stopping triggered at epoch {epoch}.")
            break

    elapsed = elapsed_before_resume + (time.time() - start_time)
    summary = {
        "best_val_iou": best_val_iou,
        "best_val_dice": max((float(r["val_dice"]) for r in history if r.get("val_dice") is not None), default=0.0),
        "final_epoch": int(history[-1]["epoch"]) if history else int(start_epoch - 1),
        "elapsed_seconds": elapsed,
        "seconds_per_epoch": elapsed / max(len(history), 1),
        "device_used": str(DEVICE),
        "strategy": strategy,
        "run_type": run_type,
        "resumed": resume_source is not None,
    }
    if resume_source is not None:
        summary["resume_source"] = resume_source
    if save_artifacts:
        save_json(run_dir / "summary.json", summary)
    return summary, history

"""=============================================================================
EVALUATION + SMOKE TEST
=============================================================================
"""

def _save_rgb_panel(image_chw: np.ndarray, pred_hw: np.ndarray, gt_hw: np.ndarray, output_path: Path, title: str) -> None:
    img = image_chw.transpose(1, 2, 0)
    img = (img - img.min()) / (img.max() - img.min() + 1e-8)
    fig, axes = plt.subplots(1, 3, figsize=(12, 4))
    axes[0].imshow(img)
    axes[0].set_title("Input")
    axes[1].imshow(pred_hw, cmap="gray", vmin=0, vmax=1)
    axes[1].set_title("Prediction")
    axes[2].imshow(gt_hw, cmap="gray", vmin=0, vmax=1)
    axes[2].set_title("Ground Truth")
    for ax in axes:
        ax.axis("off")
    fig.suptitle(title)
    fig.tight_layout()
    fig.savefig(output_path, dpi=120)
    plt.close(fig)

def evaluate_model(
    *,
    model: nn.Module,
    model_config: RuntimeModelConfig,
    bundle: DataBundle,
    run_dir: Path,
    strategy: int,
    tmax: int,
) -> tuple[dict[str, dict[str, float]], list[dict[str, Any]]]:
    amp_dtype = resolve_amp_dtype(AMP_DTYPE)
    use_amp = amp_autocast_enabled(DEVICE)
    use_channels_last = _use_channels_last_for_run(model_config)
    pred_dir = ensure_dir(run_dir / "predictions")
    pred_255_dir = ensure_dir(run_dir / "predictions_255")

    model.eval()
    per_metric = {k: [] for k in ("dice", "ppv", "sen", "iou", "biou", "hd95")}
    per_sample: list[dict[str, Any]] = []

    with torch.inference_mode():
        prefetcher = CUDAPrefetcher(bundle.test_loader, DEVICE)
        for batch in tqdm(prefetcher, total=len(bundle.test_loader), desc="Evaluating", leave=False):
            image = batch["image"]
            gt = batch["mask"]
            if use_amp and DEVICE.type == "cuda":
                image = image.to(dtype=amp_dtype)
            pred = infer_segmentation_mask(
                model,
                image,
                tmax,
                strategy=strategy,
                use_amp=use_amp,
                amp_dtype=amp_dtype,
                use_channels_last=use_channels_last,
            ).float()
            pred_np = pred.cpu().numpy().astype(np.uint8)
            gt_np = gt.cpu().numpy().astype(np.uint8)
            sample_ids = batch["sample_id"]
            for idx in range(pred_np.shape[0]):
                metrics = compute_all_metrics(pred_np[idx], gt_np[idx])
                for key, value in metrics.items():
                    per_metric[key].append(value)
                per_sample.append({"sample_id": sample_ids[idx], **metrics})
                mask_2d = pred_np[idx].squeeze()
                PILImage.fromarray(mask_2d).save(pred_dir / f"{sample_ids[idx]}.png")
                PILImage.fromarray((mask_2d * 255).astype(np.uint8)).save(pred_255_dir / f"{sample_ids[idx]}.png")

    aggregate: dict[str, dict[str, float]] = {}
    for key, values in per_metric.items():
        values_np = np.array(values, dtype=np.float32)
        aggregate[key] = {"mean": float(values_np.mean()), "std": float(values_np.std())}

    save_json(
        run_dir / "evaluation.json",
        {
            "strategy": strategy,
            "best_metric_name": BEST_CHECKPOINT_METRIC,
            "metrics": aggregate,
            "per_sample": per_sample,
        },
    )

    df_all = pd.DataFrame(per_sample)
    avg_row = {}
    for column in df_all.columns:
        avg_row[column] = df_all[column].mean() if pd.api.types.is_numeric_dtype(df_all[column]) else "AVERAGE"
    df_samples = pd.concat([df_all, pd.DataFrame([avg_row])], ignore_index=True)
    df_summary = pd.DataFrame(aggregate).T
    df_summary.index.name = "metric"
    df_low_iou = df_all[df_all["iou"] < 0.01]
    history_path = run_dir / "history.json"
    df_history = pd.DataFrame(load_json(history_path)) if history_path.exists() else None

    xlsx_path = run_dir / "evaluation_results.xlsx"
    with pd.ExcelWriter(xlsx_path, engine="openpyxl") as writer:
        df_samples.to_excel(writer, sheet_name="Per Sample", index=False)
        df_summary.to_excel(writer, sheet_name="Summary")
        if not df_low_iou.empty:
            df_low_iou.to_excel(writer, sheet_name="Low IoU Samples", index=False)
        if df_history is not None:
            df_history.to_excel(writer, sheet_name="Training History", index=False)

    csv_rows = [{"sample_id": row["sample_id"]} for row in df_low_iou.to_dict(orient="records")]
    save_json(run_dir / "evaluation_summary.json", {"mean_iou": aggregate["iou"]["mean"], "mean_dice": aggregate["dice"]["mean"]})
    pd.DataFrame(csv_rows).to_csv(run_dir / "low_iou_samples.csv", index=False)
    return aggregate, per_sample

def percent_root(percent: float) -> Path:
    return ensure_dir(RUNS_ROOT / MODEL_NAME / f"pct_{percent_label(percent)}")

def strategy_dir_name(strategy: int, model_config: RuntimeModelConfig | None = None) -> str:
    model_config = (model_config or current_model_config()).validate()
    if model_config.backbone_family == "custom_vgg":
        return f"strategy_{strategy}_custom_vgg"
    return f"strategy_{strategy}"

def strategy_root_for_percent(
    strategy: int,
    percent: float,
    model_config: RuntimeModelConfig | None = None,
) -> Path:
    return ensure_dir(percent_root(percent) / strategy_dir_name(strategy, model_config))

def final_root_for_strategy(
    strategy: int,
    percent: float,
    model_config: RuntimeModelConfig | None = None,
) -> Path:
    return ensure_dir(strategy_root_for_percent(strategy, percent, model_config) / "final")

def ensure_specific_checkpoint_scope(selector_name: str, selector_mode: str) -> None:
    if selector_mode != "specific":
        return
    # Allow STRATEGY2 checkpoints with dict mapping for dynamic per-percent selection
    if "strategy2" in selector_name.lower() and isinstance(STRATEGY2_SPECIFIC_CHECKPOINT, dict):
        return
    if len(STRATEGIES) != 1 or len(DATASET_PERCENTS) != 1:
        raise ValueError(
            f"{selector_name}=specific is only supported when exactly one strategy and one dataset percent are selected. "
            f"Got STRATEGIES={STRATEGIES} and DATASET_PERCENTS={DATASET_PERCENTS}."            
        )

def resolve_checkpoint_path(
    *,
    run_dir: Path,
    selector_mode: str,
    specific_checkpoint: str | Path | dict,
    purpose: str,
) -> Path:
    run_dir = Path(run_dir)
    if selector_mode == "latest":
        checkpoint_path = run_dir / "checkpoints" / "latest.pt"
    elif selector_mode == "best":
        checkpoint_path = run_dir / "checkpoints" / "best.pt"
    elif selector_mode == "specific":
        ensure_specific_checkpoint_scope(purpose, selector_mode)
        if not specific_checkpoint:
            raise ValueError(f"{purpose}=specific requires a non-empty specific checkpoint path.")
        checkpoint_path = Path(specific_checkpoint).expanduser().resolve()
    else:
        raise ValueError(f"Unsupported checkpoint selector mode '{selector_mode}' for {purpose}.")

    if not checkpoint_path.exists():
        raise FileNotFoundError(f"Checkpoint for {purpose} not found: {checkpoint_path}")
    return checkpoint_path

def resolve_train_resume_checkpoint_path(run_dir: Path) -> Path | None:
    if TRAIN_RESUME_MODE == "off":
        return None
    return resolve_checkpoint_path(
        run_dir=run_dir,
        selector_mode=TRAIN_RESUME_MODE,
        specific_checkpoint=TRAIN_RESUME_SPECIFIC_CHECKPOINT,
        purpose="train_resume_checkpoint",
    )

def resolve_strategy2_checkpoint_path(
    strategy: int,
    percent: float,
    model_config: RuntimeModelConfig | None = None,
) -> Path:
    if strategy not in (3, 4, 5):
        raise ValueError(f"Strategy 2 dependency checkpoint requested for unsupported strategy {strategy}.")

    # --- ADD THIS BLOCK ---
    specific_checkpoint = STRATEGY2_SPECIFIC_CHECKPOINT
    if isinstance(specific_checkpoint, dict):
        specific_checkpoint = specific_checkpoint.get(percent, "")
    # ----------------------

    checkpoint_path = resolve_checkpoint_path(
        run_dir=final_root_for_strategy(2, percent, model_config),
        selector_mode=STRATEGY2_CHECKPOINT_MODE,
        specific_checkpoint=specific_checkpoint,
        purpose="strategy2_checkpoint",
    )
    
    # Print which checkpoint is being used
    print(f"[Strategy 2 Checkpoint] Strategy {strategy} | {percent_label(percent)}% | Loading: {checkpoint_path}")
    
    return checkpoint_path

def load_required_hparams(payload: dict[str, Any], *, source: str, strategy: int, percent: float) -> dict[str, Any]:
    missing_keys = [name for name in REQUIRED_HPARAM_KEYS if name not in payload]
    if missing_keys:
        raise KeyError(
            f"Incomplete hyperparameters for strategy={strategy}, percent={percent_label(percent)}% from {source}. "
            f"Missing keys: {missing_keys}. Required keys: {REQUIRED_HPARAM_KEYS}."
        )
    return dict(payload)

def load_saved_best_params_if_optuna_off(
    strategy: int,
    percent: float,
    *,
    model_config: RuntimeModelConfig,
) -> dict[str, Any]:
    _, study_root, _ = study_paths_for(strategy, percent, model_config)
    best_params_path = study_root / "best_params.json"
    if not best_params_path.exists():
        raise FileNotFoundError(
            f"RUN_OPTUNA=False and USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF=True, but no saved best params were found "
            f"for strategy={strategy}, percent={percent_label(percent)}% at {best_params_path}."
        )
    params = load_json(best_params_path)
    params = load_required_hparams(
        params,
        source=str(best_params_path),
        strategy=strategy,
        percent=percent,
    )
    print(
        f"[Optuna Off] Using saved best parameters for strategy={strategy}, "
        f"percent={percent_label(percent)}% from {best_params_path}."
    )
    for name in REQUIRED_HPARAM_KEYS:
        print(f"  {name:20s}: {_format_hparam_value(name, params[name])}")
    return params

def load_manual_hparams_if_optuna_off(strategy: int, percent: float) -> dict[str, Any]:
    key = manual_hparams_key(strategy, percent)
    if key not in MANUAL_HPARAMS_IF_OPTUNA_OFF:
        raise KeyError(
            f"RUN_OPTUNA=False and USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF=False, but no manual hyperparameter "
            f"JSON filename was found for strategy={strategy}, percent={percent_label(percent)}% under key '{key}'. "
            f"Required keys: {REQUIRED_HPARAM_KEYS}."
        )
    manual_filename = MANUAL_HPARAMS_IF_OPTUNA_OFF[key]
    manual_path = (HARD_CODED_PARAM_DIR / manual_filename).resolve()
    if not manual_path.exists():
        raise FileNotFoundError(
            f"RUN_OPTUNA=False and USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF=False, but the manual hyperparameter "
            f"JSON file for strategy={strategy}, percent={percent_label(percent)}% was not found at {manual_path}. "
            f"Configured key='{key}', filename='{manual_filename}'."
        )
    params = load_json(manual_path)
    params = load_required_hparams(
        params,
        source=str(manual_path),
        strategy=strategy,
        percent=percent,
    )
    print(
        f"[Optuna Off] Using manual hyperparameters for strategy={strategy}, "
        f"percent={percent_label(percent)}% from {manual_path}."
    )
    for name in REQUIRED_HPARAM_KEYS:
        print(f"  {name:20s}: {_format_hparam_value(name, params[name])}")
    return params

def resolve_job_params(
    strategy: int,
    percent: float,
    bundle: DataBundle,
    *,
    model_config: RuntimeModelConfig,
    strategy2_checkpoint_path: str | Path | None = None,
) -> dict[str, Any]:
    if RUN_OPTUNA:
        banner(f"OPTUNA STUDY | Strategy {strategy} | {percent_label(percent)}%")
        return run_study(
            strategy,
            bundle,
            model_config=model_config,
            strategy2_checkpoint_path=strategy2_checkpoint_path,
        )

    if USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF:
        return load_saved_best_params_if_optuna_off(strategy, percent, model_config=model_config)

    return load_manual_hparams_if_optuna_off(strategy, percent)

def read_run_config_for_eval(run_dir: Path, checkpoint_path: Path) -> dict[str, Any]:
    run_config_path = Path(run_dir) / "run_config.json"
    if run_config_path.exists():
        return load_json(run_config_path)
    ckpt = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
    return checkpoint_run_config_payload(ckpt)

def run_evaluation_for_run(
    *,
    strategy: int,
    percent: float,
    bundle: DataBundle,
    run_dir: Path,
    strategy2_checkpoint_path: str | Path | None = None,
) -> tuple[dict[str, dict[str, float]], list[dict[str, Any]]]:
    checkpoint_path = resolve_checkpoint_path(
        run_dir=run_dir,
        selector_mode=EVAL_CHECKPOINT_MODE,
        specific_checkpoint=EVAL_SPECIFIC_CHECKPOINT,
        purpose="evaluation_checkpoint",
    )
    effective_run_dir = Path(run_dir)
    if not (effective_run_dir / "run_config.json").exists() and checkpoint_path.parent.name == "checkpoints":
        effective_run_dir = checkpoint_path.parent.parent
    print(f"[Evaluation] Strategy {strategy} | {percent_label(percent)}% | checkpoint={checkpoint_path}")
    runtime_config = read_run_config_for_eval(effective_run_dir, checkpoint_path)
    set_current_job_params(runtime_config)
    model_config = RuntimeModelConfig.from_payload(runtime_config).validate()
    if strategy in (4, 5):
        if runtime_config.get("strategy2_checkpoint_path"):
            strategy2_checkpoint_path = runtime_config["strategy2_checkpoint_path"]
        else:
            strategy2_checkpoint_path = resolve_strategy2_checkpoint_path(strategy, percent, model_config)
    elif strategy == 3 and runtime_config.get("strategy2_checkpoint_path"):
        strategy2_checkpoint_path = runtime_config["strategy2_checkpoint_path"]
    dropout_p = float(runtime_config.get("dropout_p", DEFAULT_DROPOUT_P))
    tmax = int(runtime_config.get("tmax", DEFAULT_TMAX))
    eval_model, _description, _compiled = build_model(
        strategy,
        dropout_p,
        model_config=model_config,
        strategy2_checkpoint_path=strategy2_checkpoint_path,
    )
    load_checkpoint(
        checkpoint_path,
        model=eval_model,
        device=DEVICE,
    )
    aggregate, per_sample = evaluate_model(
        model=eval_model,
        model_config=model_config,
        bundle=bundle,
        run_dir=effective_run_dir,
        strategy=strategy,
        tmax=tmax,
    )
    evaluation_json_path = effective_run_dir / "evaluation.json"
    evaluation_payload = load_json(evaluation_json_path)
    evaluation_payload["checkpoint_mode"] = EVAL_CHECKPOINT_MODE
    evaluation_payload["checkpoint_path"] = str(checkpoint_path)
    save_json(evaluation_json_path, evaluation_payload)
    del eval_model
    run_cuda_cleanup(context=f"evaluation strategy={strategy} percent={percent_label(percent)}%")
    return aggregate, per_sample

def run_smoke_test(
    *,
    strategy: int,
    model_config: RuntimeModelConfig,
    bundle: DataBundle,
    smoke_root: Path,
    strategy2_checkpoint_path: str | Path | None = None,
) -> None:
    banner(f"PRE-TRAINING SMOKE TEST | Strategy {strategy} | {percent_label(bundle.percent)}%")
    smoke_root = ensure_dir(smoke_root)
    sample = bundle.test_ds[SMOKE_TEST_SAMPLE_INDEX]
    image = sample["image"].unsqueeze(0).to(DEVICE)
    raw_image = sample["image"].numpy()
    raw_gt = sample["mask"].squeeze(0).numpy()
    amp_dtype = resolve_amp_dtype(AMP_DTYPE)
    use_amp = amp_autocast_enabled(DEVICE)
    if use_amp and DEVICE.type == "cuda":
        image = image.to(dtype=amp_dtype)

    model, description, compiled = build_model(
        strategy,
        DEFAULT_DROPOUT_P,
        model_config=model_config,
        strategy2_checkpoint_path=strategy2_checkpoint_path,
    )
    print_model_parameter_summary(
        model=model,
        description=f"{description} | Smoke Test",
        strategy=strategy,
        model_config=model_config,
        dropout_p=DEFAULT_DROPOUT_P,
        amp_dtype=amp_dtype,
        compiled=compiled,
    )
    pred = infer_segmentation_mask(
        model,
        image,
        DEFAULT_TMAX,
        strategy=strategy,
        use_amp=use_amp,
        amp_dtype=amp_dtype,
        use_channels_last=_use_channels_last_for_run(model_config),
    ).float()
    pred_np = pred[0, 0].detach().cpu().numpy()
    panel_path = smoke_root / "smoke_panel.png"
    raw_mask_path = smoke_root / "smoke_prediction.png"
    _save_rgb_panel(raw_image, pred_np, raw_gt, panel_path, f"Smoke Test | Strategy {strategy}")
    PILImage.fromarray((pred_np * 255).astype(np.uint8)).save(raw_mask_path)
    del model
    run_cuda_cleanup(context=f"smoke strategy={strategy} percent={percent_label(bundle.percent)}%")
    print(
        f"[Smoke Test] Strategy {strategy} passed. "
        f"Saved panel to {panel_path.name} and mask to {raw_mask_path.name}."
    )

"""=============================================================================
OVERFIT TEST
=============================================================================
"""

OVERFIT_HISTORY_KEYS = (
    "dice",
    "iou",
    "loss",
    "reward",
    "actor_loss",
    "critic_loss",
    "ce_loss",
    "dice_loss",
    "entropy",
    "grad_norm",
    "action_dist",
    "reward_pos_pct",
    "pred_fg_pct",
    "gt_fg_pct",
)

def empty_overfit_history() -> dict[str, list[Any]]:
    return {key: [] for key in OVERFIT_HISTORY_KEYS}

def load_overfit_history_for_resume(history_path: Path, checkpoint_payload: dict[str, Any]) -> dict[str, list[Any]]:
    history = empty_overfit_history()
    checkpoint_epoch = int(checkpoint_payload.get("epoch", 0))
    if history_path.exists():
        payload = load_json(history_path)
        if not isinstance(payload, dict):
            raise RuntimeError(f"Expected dict history at {history_path}, found {type(payload).__name__}.")
        for key in OVERFIT_HISTORY_KEYS:
            values = payload.get(key, [])
            if isinstance(values, list):
                history[key] = list(values[:checkpoint_epoch])
        return history

    epoch_metrics = checkpoint_payload.get("epoch_metrics", {})
    if isinstance(epoch_metrics, dict):
        for key in OVERFIT_HISTORY_KEYS:
            if key in epoch_metrics:
                history[key].append(epoch_metrics[key])
    return history

def _grad_diagnostics(model: nn.Module) -> dict[str, Any]:
    raw = _unwrap_compiled(model)
    groups: dict[str, list[float]] = {}
    total_sq = 0.0
    n_nan = 0
    n_inf = 0
    n_zero = 0
    n_total_params = 0

    for name, param in raw.named_parameters():
        if not param.requires_grad:
            continue
        n_total_params += 1
        if param.grad is None:
            n_zero += 1
            continue
        grad_norm = float(param.grad.data.norm(2).item())
        if math.isnan(grad_norm):
            n_nan += 1
            continue
        if math.isinf(grad_norm):
            n_inf += 1
            continue
        total_sq += grad_norm ** 2
        group_name = name.split(".", 1)[0]
        groups.setdefault(group_name, []).append(grad_norm)

    group_stats: dict[str, dict[str, float | int]] = {}
    for group_name, norms in groups.items():
        group_stats[group_name] = {
            "min": min(norms),
            "max": max(norms),
            "mean": sum(norms) / len(norms),
            "count": len(norms),
        }
    return {
        "global_norm": total_sq ** 0.5,
        "groups": group_stats,
        "n_nan": n_nan,
        "n_inf": n_inf,
        "n_zero_grad": n_zero,
        "n_total": n_total_params,
    }

def _param_diagnostics(model: nn.Module, prev_params: dict[str, torch.Tensor] | None = None) -> dict[str, dict[str, float]]:
    raw = _unwrap_compiled(model)
    info: dict[str, dict[str, list[float]]] = {}
    for name, param in raw.named_parameters():
        if not param.requires_grad:
            continue
        param_norm = float(param.data.norm(2).item())
        group_name = name.split(".", 1)[0]
        entry = info.setdefault(group_name, {"norms": [], "update_ratios": []})
        entry["norms"].append(param_norm)
        if prev_params is not None and name in prev_params:
            delta = float((param.data - prev_params[name]).norm(2).item())
            entry["update_ratios"].append(delta / max(param_norm, 1e-12))

    summary: dict[str, dict[str, float]] = {}
    for group_name, values in info.items():
        norms = values["norms"]
        ratios = values["update_ratios"]
        summary[group_name] = {
            "p_min": min(norms),
            "p_max": max(norms),
            "p_mean": sum(norms) / len(norms),
        }
        if ratios:
            summary[group_name]["ur_min"] = min(ratios)
            summary[group_name]["ur_max"] = max(ratios)
            summary[group_name]["ur_mean"] = sum(ratios) / len(ratios)
    return summary

def _snapshot_params(model: nn.Module) -> dict[str, torch.Tensor]:
    raw = _unwrap_compiled(model)
    return {
        name: param.data.detach().clone()
        for name, param in raw.named_parameters()
        if param.requires_grad
    }

def _action_distribution(
    model: nn.Module,
    image: torch.Tensor,
    seg: torch.Tensor,
    tmax: int,
    use_amp: bool,
    amp_dtype: torch.dtype,
    *,
    strategy: int | None = None,
) -> tuple[list[dict[int, float]], torch.Tensor]:
    distributions: list[dict[int, float]] = []
    refinement_context: dict[str, torch.Tensor] | None = None
    if strategy in (3, 4) and _uses_refinement_runtime(model, strategy=strategy):
        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
            refinement_context = model.prepare_refinement_context(image)
        seg = seg.float()
    for _step in range(tmax):
        if refinement_context is not None:
            with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                state, _ = model.forward_refinement_state(
                    refinement_context["base_features"],
                    seg,
                    refinement_context["decoder_prob"],
                )
                policy_logits, _ = model.forward_from_state(state)
                actions, _, _ = sample_actions(policy_logits, stochastic=False)
            seg = apply_actions(seg, actions, soft_update_step=DEFAULT_MASK_UPDATE_STEP).to(dtype=seg.dtype)
        else:
            with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                masked = image * seg
                policy_logits = model.forward_policy_only(masked)
                actions, _, _ = sample_actions(policy_logits, stochastic=False)
            seg = apply_actions(seg, actions, num_actions=policy_logits.shape[1]).to(dtype=seg.dtype)
        total_pixels = max(actions.numel(), 1)
        step_dist: dict[int, float] = {}
        action_count = int(policy_logits.shape[1])
        for action_idx in range(action_count):
            step_dist[action_idx] = float((actions == action_idx).sum().item()) / total_pixels * 100.0
        distributions.append(step_dist)
    if refinement_context is not None:
        return distributions, threshold_binary_mask(seg.float()).float()
    return distributions, seg

def _numerical_health_check(outputs_dict: dict[str, Any], prefix: str = "") -> list[str]:
    alerts: list[str] = []
    for name, value in outputs_dict.items():
        if value is None:
            continue
        if isinstance(value, (int, float)):
            if math.isnan(value):
                alerts.append(f"{prefix}{name} = NaN")
            elif math.isinf(value):
                alerts.append(f"{prefix}{name} = Inf")
            continue
        if torch.is_tensor(value):
            if torch.isnan(value).any():
                alerts.append(f"{prefix}{name} contains NaN")
            if torch.isinf(value).any():
                alerts.append(f"{prefix}{name} contains Inf")
    return alerts

def _batch_binary_metrics(pred: torch.Tensor, gt: torch.Tensor) -> tuple[list[float], list[float]]:
    pred_np = pred.detach().cpu().numpy().astype(np.uint8)
    gt_np = gt.detach().cpu().numpy().astype(np.uint8)
    dices: list[float] = []
    ious: list[float] = []
    for idx in range(pred_np.shape[0]):
        metrics = compute_all_metrics(pred_np[idx], gt_np[idx])
        dices.append(float(metrics["dice"]))
        ious.append(float(metrics["iou"]))
    return dices, ious

def run_overfit_test(
    *,
    strategy: int,
    model_config: RuntimeModelConfig,
    bundle: DataBundle,
    overfit_root: Path,
    strategy2_checkpoint_path: str | Path | None = None,
) -> dict[str, Any]:
    amp_dtype = resolve_amp_dtype(AMP_DTYPE)
    use_amp = amp_autocast_enabled(DEVICE)
    use_channels_last = _use_channels_last_for_run(model_config)
    overfit_root = ensure_dir(overfit_root)
    ckpt_dir = ensure_dir(overfit_root / "checkpoints")
    history_path = checkpoint_history_path(overfit_root, "overfit")
    run_config = {
        "project_dir": str(PROJECT_DIR),
        "data_root": str(DATA_ROOT),
        "run_type": "overfit",
        "strategy": strategy,
        "dataset_percent": bundle.percent,
        "dataset_name": bundle.split_payload["dataset_name"],
        "dataset_split_policy": bundle.split_payload["dataset_split_policy"],
        "dataset_splits_path": bundle.split_payload["dataset_splits_path"],
        "split_type": bundle.split_payload["split_type"],
        "train_subset_key": bundle.split_payload["train_subset_key"],
        "max_epochs": OVERFIT_N_EPOCHS,
        "head_lr": OVERFIT_HEAD_LR,
        "encoder_lr": OVERFIT_ENCODER_LR,
        "weight_decay": DEFAULT_WEIGHT_DECAY,
        "dropout_p": DEFAULT_DROPOUT_P,
        "tmax": DEFAULT_TMAX,
        "entropy_lr": DEFAULT_ENTROPY_LR,
        "gamma": DEFAULT_GAMMA,
        "grad_clip_norm": DEFAULT_GRAD_CLIP_NORM,
        "train_resume_mode": TRAIN_RESUME_MODE,
        "train_resume_specific_checkpoint": TRAIN_RESUME_SPECIFIC_CHECKPOINT,
    }
    run_config.update(model_config.to_payload())
    if strategy2_checkpoint_path is not None:
        run_config["strategy2_checkpoint_path"] = str(Path(strategy2_checkpoint_path))
    save_json(overfit_root / "run_config.json", run_config)
    set_current_job_params(run_config)

    banner(f"OVERFIT TEST | Strategy {strategy} | {percent_label(bundle.percent)}%")
    print(
        f"[Overfit] Fixed batches={OVERFIT_N_BATCHES}, epochs={OVERFIT_N_EPOCHS}, "
        f"head_lr={OVERFIT_HEAD_LR:.2e}, encoder_lr={OVERFIT_ENCODER_LR:.2e}"
    )

    fixed_batches: list[dict[str, Any]] = []
    for batch_index, batch in enumerate(bundle.train_loader):
        fixed_batches.append(to_device(batch, DEVICE))
        if batch_index + 1 >= OVERFIT_N_BATCHES:
            break
    if not fixed_batches:
        raise RuntimeError("Overfit test could not collect any training batches.")
    if len(fixed_batches) < OVERFIT_N_BATCHES:
        print(f"[Overfit] Warning: only {len(fixed_batches)} train batch(es) available.")

    model, description, compiled = build_model(
        strategy,
        DEFAULT_DROPOUT_P,
        model_config=model_config,
        strategy2_checkpoint_path=strategy2_checkpoint_path,
    )
    resume_checkpoint_path = resolve_train_resume_checkpoint_path(overfit_root)
    if resume_checkpoint_path is not None and strategy in (3, 4):
        _configure_model_from_checkpoint_path(model, resume_checkpoint_path)
    print_model_parameter_summary(
        model=model,
        description=f"{description} | Overfit Test",
        strategy=strategy,
        model_config=model_config,
        dropout_p=DEFAULT_DROPOUT_P,
        amp_dtype=amp_dtype,
        compiled=compiled,
    )

    optimizer = make_optimizer(
        model,
        strategy,
        head_lr=OVERFIT_HEAD_LR,
        encoder_lr=OVERFIT_ENCODER_LR,
        weight_decay=DEFAULT_WEIGHT_DECAY,
    )
    scaler = make_grad_scaler(enabled=use_amp, amp_dtype=amp_dtype, device=DEVICE)
    log_alpha: torch.Tensor | None = None
    alpha_optimizer: Adam | None = None
    target_entropy = DEFAULT_ENTROPY_TARGET_RATIO * math.log(max(_model_policy_action_count(model) or NUM_ACTIONS, 2))
    if strategy != 2:
        log_alpha = torch.tensor(
            math.log(max(DEFAULT_ENTROPY_ALPHA_INIT, 1e-8)),
            dtype=torch.float32,
            device=DEVICE,
            requires_grad=True,
        )
        alpha_optimizer = Adam([log_alpha], lr=DEFAULT_ENTROPY_LR)

    history = empty_overfit_history()
    prev_loss: float | None = None
    best_dice = -1.0
    elapsed_before_resume = 0.0
    start_epoch = 1
    resume_source: dict[str, Any] | None = None
    if resume_checkpoint_path is not None:
        checkpoint_payload = load_checkpoint(
            resume_checkpoint_path,
            model=model,
            optimizer=optimizer,
            scheduler=None,
            scaler=scaler,
            device=DEVICE,
            log_alpha=log_alpha,
            alpha_optimizer=alpha_optimizer,
            expected_run_type="overfit",
            require_run_metadata=True,
        )
        validate_resume_checkpoint_identity(
            run_config,
            checkpoint_run_config_payload(checkpoint_payload),
            checkpoint_path=resume_checkpoint_path,
        )
        start_epoch = int(checkpoint_payload["epoch"]) + 1
        best_dice = float(checkpoint_payload["best_metric_value"])
        elapsed_before_resume = float(checkpoint_payload.get("elapsed_seconds", 0.0))
        history = load_overfit_history_for_resume(history_path, checkpoint_payload)
        resume_source = {
            "checkpoint_path": str(Path(resume_checkpoint_path).resolve()),
            "checkpoint_epoch": int(checkpoint_payload["epoch"]),
            "checkpoint_run_type": checkpoint_payload.get("run_type", "overfit"),
        }
        print(
            f"[Resume] overfit run continuing from {resume_checkpoint_path} "
            f"at epoch {start_epoch}/{OVERFIT_N_EPOCHS}."
        )
    if history["loss"]:
        prev_loss = float(history["loss"][-1])
    prev_params = _snapshot_params(model)
    start_time = time.time()

    for epoch in range(start_epoch, OVERFIT_N_EPOCHS + 1):
        full_dump = epoch <= 5 or epoch % max(OVERFIT_PRINT_EVERY, 1) == 0 or epoch == OVERFIT_N_EPOCHS
        epoch_losses: list[float] = []
        epoch_dices: list[float] = []
        epoch_ious: list[float] = []
        epoch_rewards: list[float] = []
        epoch_actor: list[float] = []
        epoch_critic: list[float] = []
        epoch_ce: list[float] = []
        epoch_dice_losses: list[float] = []
        epoch_entropy: list[float] = []
        epoch_grad_norms: list[float] = []
        epoch_action_dist: list[list[dict[int, float]]] = []
        epoch_reward_pos_pct: list[float] = []
        epoch_pred_fg_pct: list[float] = []
        epoch_gt_fg_pct: list[float] = []
        epoch_alerts: list[str] = []

        for batch in fixed_batches:
            if strategy == 2:
                metrics = train_step_supervised(
                    model,
                    batch,
                    optimizer,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    use_channels_last=use_channels_last,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    ce_weight=DEFAULT_CE_WEIGHT,
                    dice_weight=DEFAULT_DICE_WEIGHT,
                )
            elif strategy in (3, 4):
                assert log_alpha is not None and alpha_optimizer is not None
                metrics = train_step_strategy3(
                    model,
                    batch,
                    optimizer,
                    gamma=DEFAULT_GAMMA,
                    tmax=DEFAULT_TMAX,
                    critic_loss_weight=DEFAULT_CRITIC_LOSS_WEIGHT,
                    log_alpha=log_alpha,
                    alpha_optimizer=alpha_optimizer,
                    target_entropy=target_entropy,
                    ce_weight=DEFAULT_CE_WEIGHT,
                    dice_weight=DEFAULT_DICE_WEIGHT,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    stepwise_backward=STEPWISE_BACKWARD,
                    use_channels_last=use_channels_last,
                )
            else:
                assert log_alpha is not None and alpha_optimizer is not None
                metrics = train_step(
                    model,
                    batch,
                    optimizer,
                    gamma=DEFAULT_GAMMA,
                    tmax=DEFAULT_TMAX,
                    critic_loss_weight=DEFAULT_CRITIC_LOSS_WEIGHT,
                    log_alpha=log_alpha,
                    alpha_optimizer=alpha_optimizer,
                    target_entropy=target_entropy,
                    ce_weight=DEFAULT_CE_WEIGHT,
                    dice_weight=DEFAULT_DICE_WEIGHT,
                    grad_clip_norm=DEFAULT_GRAD_CLIP_NORM,
                    scaler=scaler,
                    use_amp=use_amp,
                    amp_dtype=amp_dtype,
                    stepwise_backward=STEPWISE_BACKWARD,
                    use_channels_last=use_channels_last,
                )

            epoch_losses.append(float(metrics["loss"]))
            epoch_rewards.append(float(metrics["mean_reward"]))
            epoch_actor.append(float(metrics["actor_loss"]))
            epoch_critic.append(float(metrics["critic_loss"]))
            epoch_ce.append(float(metrics["ce_loss"]))
            epoch_dice_losses.append(float(metrics["dice_loss"]))
            epoch_entropy.append(float(metrics["entropy"]))
            epoch_grad_norms.append(float(metrics["grad_norm"]))
            epoch_alerts.extend(
                _numerical_health_check(
                    {
                        "loss": metrics["loss"],
                        "actor_loss": metrics["actor_loss"],
                        "critic_loss": metrics["critic_loss"],
                        "reward": metrics["mean_reward"],
                        "entropy": metrics["entropy"],
                        "grad_norm": metrics["grad_norm"],
                        "ce_loss": metrics["ce_loss"],
                        "dice_loss": metrics["dice_loss"],
                    },
                    prefix="train:",
                )
            )

            model.eval()
            with torch.inference_mode():
                image = batch["image"]
                gt_mask = batch["mask"].float()
                if use_channels_last and image.ndim == 4 and image.device.type == "cuda":
                    image = image.contiguous(memory_format=torch.channels_last)
                if use_amp and DEVICE.type == "cuda":
                    image = image.to(dtype=amp_dtype)

                if strategy == 2:
                    with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                        logits = model(image)
                    pred = threshold_binary_mask(torch.sigmoid(logits)).float()
                else:
                    refinement_runtime = _uses_refinement_runtime(model, strategy=strategy)
                    if strategy in (3, 4) and refinement_runtime:
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            init_mask = model.prepare_refinement_context(image)["decoder_prob"].float()
                    else:
                        init_mask = torch.ones(
                            image.shape[0],
                            1,
                            image.shape[2],
                            image.shape[3],
                            device=image.device,
                            dtype=image.dtype,
                        )
                        if strategy in (3, 4):
                            with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                                init_mask = threshold_binary_mask(torch.sigmoid(model.forward_decoder(image))).float()
                    action_dist, pred = _action_distribution(
                        model,
                        image,
                        init_mask,
                        DEFAULT_TMAX,
                        use_amp,
                        amp_dtype,
                        strategy=strategy,
                    )
                    epoch_action_dist.append(action_dist)

                    if strategy in (3, 4) and refinement_runtime:
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            refinement_context = model.prepare_refinement_context(image)
                            soft_init_mask = refinement_context["decoder_prob"].float()
                            state_t, _ = model.forward_refinement_state(
                                refinement_context["base_features"],
                                soft_init_mask,
                                refinement_context["decoder_prob"],
                            )
                            policy_logits, _ = model.forward_from_state(state_t)
                            first_actions, _, _ = sample_actions(policy_logits, stochastic=False)
                        first_seg = apply_actions(soft_init_mask, first_actions, soft_update_step=DEFAULT_MASK_UPDATE_STEP, num_actions=policy_logits.shape[1])
                        reward_map = compute_refinement_reward(
                            soft_init_mask,
                            first_seg,
                            gt_mask.float(),
                            decoder_prior=refinement_context["decoder_prob"].detach(),
                        )
                    else:
                        with autocast_ctx(enabled=use_amp, device=image.device, amp_dtype=amp_dtype):
                            masked = image * init_mask
                            policy_logits = model.forward_policy_only(masked)
                            first_actions, _, _ = sample_actions(policy_logits, stochastic=False)
                        first_seg = apply_actions(init_mask, first_actions, num_actions=policy_logits.shape[1])
                        reward_map = (init_mask - gt_mask).pow(2) - (first_seg - gt_mask).pow(2)
                    epoch_reward_pos_pct.append(float((reward_map > 0).float().mean().item() * 100.0))

                pred_fg_pct = float(pred.sum().item()) / max(pred.numel(), 1) * 100.0
                gt_fg_pct = float(gt_mask.sum().item()) / max(gt_mask.numel(), 1) * 100.0
                epoch_pred_fg_pct.append(pred_fg_pct)
                epoch_gt_fg_pct.append(gt_fg_pct)
                dice_values, iou_values = _batch_binary_metrics(pred.float(), gt_mask.float())
                epoch_dices.extend(dice_values)
                epoch_ious.extend(iou_values)
                epoch_alerts.extend(
                    _numerical_health_check(
                        {"pred": pred, "gt_mask": gt_mask},
                        prefix="eval:",
                    )
                )
            model.train()

        grad_stats = _grad_diagnostics(model)
        param_stats = _param_diagnostics(model, prev_params)
        prev_params = _snapshot_params(model)

        avg_loss = float(np.mean(epoch_losses)) if epoch_losses else 0.0
        avg_dice = float(np.mean(epoch_dices)) if epoch_dices else 0.0
        avg_iou = float(np.mean(epoch_ious)) if epoch_ious else 0.0
        avg_reward = float(np.mean(epoch_rewards)) if epoch_rewards else 0.0
        avg_actor = float(np.mean(epoch_actor)) if epoch_actor else 0.0
        avg_critic = float(np.mean(epoch_critic)) if epoch_critic else 0.0
        avg_ce = float(np.mean(epoch_ce)) if epoch_ce else 0.0
        avg_dice_loss = float(np.mean(epoch_dice_losses)) if epoch_dice_losses else 0.0
        avg_entropy = float(np.mean(epoch_entropy)) if epoch_entropy else 0.0
        avg_grad_norm = float(np.mean(epoch_grad_norms)) if epoch_grad_norms else 0.0
        avg_reward_pos = float(np.mean(epoch_reward_pos_pct)) if epoch_reward_pos_pct else 0.0
        avg_pred_fg = float(np.mean(epoch_pred_fg_pct)) if epoch_pred_fg_pct else 0.0
        avg_gt_fg = float(np.mean(epoch_gt_fg_pct)) if epoch_gt_fg_pct else 0.0

        avg_action_dist: list[dict[int, float]] = []
        if epoch_action_dist:
            for step_idx in range(DEFAULT_TMAX):
                step_dist: dict[int, float] = {}
                action_indices = sorted({action_idx for run_dist in epoch_action_dist if step_idx < len(run_dist) for action_idx in run_dist[step_idx]})
                for action_idx in action_indices:
                    values = [
                        run_dist[step_idx][action_idx]
                        for run_dist in epoch_action_dist
                        if step_idx < len(run_dist)
                    ]
                    step_dist[action_idx] = float(np.mean(values)) if values else 0.0
                avg_action_dist.append(step_dist)

        history["dice"].append(avg_dice)
        history["iou"].append(avg_iou)
        history["loss"].append(avg_loss)
        history["reward"].append(avg_reward)
        history["actor_loss"].append(avg_actor)
        history["critic_loss"].append(avg_critic)
        history["ce_loss"].append(avg_ce)
        history["dice_loss"].append(avg_dice_loss)
        history["entropy"].append(avg_entropy)
        history["grad_norm"].append(avg_grad_norm)
        history["action_dist"].append(avg_action_dist)
        history["reward_pos_pct"].append(avg_reward_pos)
        history["pred_fg_pct"].append(avg_pred_fg)
        history["gt_fg_pct"].append(avg_gt_fg)
        save_json(history_path, history)

        loss_delta = avg_loss - prev_loss if prev_loss is not None else 0.0
        prev_loss = avg_loss
        if epoch_alerts:
            print(f"[Overfit][Epoch {epoch}] Numerical alerts: {' | '.join(epoch_alerts)}")

        if full_dump:
            current_alpha = float(log_alpha.exp().detach().item()) if log_alpha is not None else 0.0
            print(
                f"[Overfit][Epoch {epoch:03d}] loss={avg_loss:.6f} (delta={loss_delta:+.6f}) "
                f"dice={avg_dice:.4f} iou={avg_iou:.4f} reward={avg_reward:+.6f} "
                f"entropy={avg_entropy:.6f} alpha={current_alpha:.4f}"
            )
            print(
                f"[Overfit][Epoch {epoch:03d}] ce={avg_ce:.6f} dice_l={avg_dice_loss:.6f} "
                f"grad_norm={avg_grad_norm:.6f} global_grad={grad_stats['global_norm']:.6f}"
            )
            if avg_action_dist:
                first = avg_action_dist[0]
                last = avg_action_dist[-1]
                print(
                    f"[Overfit][Epoch {epoch:03d}] action step0={first} step_last={last} "
                    f"reward_pos={avg_reward_pos:.2f}%"
                )
            print(
                f"[Overfit][Epoch {epoch:03d}] pred_fg={avg_pred_fg:.2f}% gt_fg={avg_gt_fg:.2f}% "
                f"param_groups={list(param_stats.keys())}"
            )
        else:
            print(
                f"[Overfit][Epoch {epoch:03d}] loss={avg_loss:.6f} dice={avg_dice:.4f} "
                f"iou={avg_iou:.4f} reward={avg_reward:+.6f}"
            )

        row = {
            "epoch": epoch,
            "dice": avg_dice,
            "iou": avg_iou,
            "loss": avg_loss,
            "reward": avg_reward,
            "actor_loss": avg_actor,
            "critic_loss": avg_critic,
            "ce_loss": avg_ce,
            "dice_loss": avg_dice_loss,
            "entropy": avg_entropy,
            "grad_norm": avg_grad_norm,
            "action_dist": avg_action_dist,
            "reward_pos_pct": avg_reward_pos,
            "pred_fg_pct": avg_pred_fg,
            "gt_fg_pct": avg_gt_fg,
        }
        if avg_dice > best_dice:
            best_dice = avg_dice
            save_checkpoint(
                ckpt_dir / "best.pt",
                run_type="overfit",
                model=model,
                optimizer=optimizer,
                scheduler=None,
                scaler=scaler,
                epoch=epoch,
                best_metric_value=best_dice,
                best_metric_name="overfit_dice",
                run_config=run_config,
                epoch_metrics=row,
                patience_counter=0,
                elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                log_alpha=log_alpha,
                alpha_optimizer=alpha_optimizer,
                resume_source=resume_source,
            )

        if SAVE_LATEST_EVERY_EPOCH:
            save_checkpoint(
                ckpt_dir / "latest.pt",
                run_type="overfit",
                model=model,
                optimizer=optimizer,
                scheduler=None,
                scaler=scaler,
                epoch=epoch,
                best_metric_value=best_dice,
                best_metric_name="overfit_dice",
                run_config=run_config,
                epoch_metrics=row,
                patience_counter=0,
                elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                log_alpha=log_alpha,
                alpha_optimizer=alpha_optimizer,
                resume_source=resume_source,
            )
        if CHECKPOINT_EVERY_N_EPOCHS > 0 and epoch % CHECKPOINT_EVERY_N_EPOCHS == 0:
            save_checkpoint(
                ckpt_dir / f"epoch_{epoch:04d}.pt",
                run_type="overfit",
                model=model,
                optimizer=optimizer,
                scheduler=None,
                scaler=scaler,
                epoch=epoch,
                best_metric_value=best_dice,
                best_metric_name="overfit_dice",
                run_config=run_config,
                epoch_metrics=row,
                patience_counter=0,
                elapsed_seconds=elapsed_before_resume + (time.time() - start_time),
                log_alpha=log_alpha,
                alpha_optimizer=alpha_optimizer,
                resume_source=resume_source,
            )

    peak_dice = max(history["dice"]) if history["dice"] else 0.0
    final_dice = history["dice"][-1] if history["dice"] else 0.0
    summary = {
        "run_type": "overfit",
        "strategy": strategy,
        "peak_dice": peak_dice,
        "final_dice": final_dice,
        "description": description,
        "resumed": resume_source is not None,
        "elapsed_seconds": elapsed_before_resume + (time.time() - start_time),
        "final_epoch": max(len(history["dice"]), start_epoch - 1),
    }
    if resume_source is not None:
        summary["resume_source"] = resume_source
    save_json(overfit_root / "summary.json", summary)
    print(
        f"[Overfit] Strategy {strategy} | {percent_label(bundle.percent)}% | "
        f"peak_dice={peak_dice:.4f}, final_dice={final_dice:.4f}"
    )

    del model
    run_cuda_cleanup(context=f"overfit strategy={strategy} percent={percent_label(bundle.percent)}%")
    return {**summary, "history": history}

def run_configured_overfit_tests(
    bundles: dict[float, DataBundle],
    *,
    model_config: RuntimeModelConfig,
) -> None:
    banner("OVERFIT TEST MODE")
    for percent in DATASET_PERCENTS:
        bundle = bundles[percent]
        for strategy in STRATEGIES:
            strategy2_checkpoint_path = None
            if strategy in (4, 5) or (strategy == 3 and STRATEGY3_BOOTSTRAP_FROM_STRATEGY2):
                strategy2_checkpoint_path = resolve_strategy2_checkpoint_path(strategy, percent, model_config)
            run_overfit_test(
                strategy=strategy,
                model_config=model_config,
                bundle=bundle,
                overfit_root=strategy_root_for_percent(strategy, percent, model_config) / "overfit_test",
                strategy2_checkpoint_path=strategy2_checkpoint_path,
            )

"""=============================================================================
OPTUNA + ORCHESTRATION
=============================================================================
"""

def strategy_epochs(strategy: int) -> int:
    if strategy == 1:
        return STRATEGY_1_MAX_EPOCHS
    if strategy == 2:
        return STRATEGY_2_MAX_EPOCHS
    if strategy == 3:
        return STRATEGY_3_MAX_EPOCHS
    if strategy == 4:
        return STRATEGY_4_MAX_EPOCHS
    if strategy == 5:
        return STRATEGY_5_MAX_EPOCHS
    raise ValueError(f"Unsupported strategy for epoch selection: {strategy}")

def suggest_hyperparameters(trial: optuna.trial.Trial, strategy: int) -> dict[str, Any]:
    head_lr = trial.suggest_float("head_lr", HEAD_LR_RANGE[0], HEAD_LR_RANGE[1], log=True)
    encoder_lr = trial.suggest_float("encoder_lr", ENCODER_LR_RANGE[0], min(ENCODER_LR_RANGE[1], head_lr), log=True)
    weight_decay = trial.suggest_float("weight_decay", WEIGHT_DECAY_RANGE[0], WEIGHT_DECAY_RANGE[1], log=True)
    dropout_p = trial.suggest_float("dropout_p", DROPOUT_P_RANGE[0], DROPOUT_P_RANGE[1])
    params = {
        "head_lr": head_lr,
        "encoder_lr": encoder_lr,
        "weight_decay": weight_decay,
        "dropout_p": dropout_p,
        "tmax": DEFAULT_TMAX,
        "entropy_lr": DEFAULT_ENTROPY_LR,
    }
    if strategy != 2:
        params["tmax"] = trial.suggest_int("tmax", TMAX_RANGE[0], TMAX_RANGE[1])
        params["entropy_lr"] = trial.suggest_float("entropy_lr", ENTROPY_LR_RANGE[0], ENTROPY_LR_RANGE[1], log=True)
    return params

def _format_hparam_value(key: str, value: Any) -> str:
    if isinstance(value, float):
        if key in {"head_lr", "encoder_lr", "entropy_lr"}:
            return f"{value:.3e}"
        return f"{value:.6g}"
    return str(value)

def log_optuna_trial_start(
    *,
    study_name: str,
    strategy: int,
    bundle: DataBundle,
    trial: optuna.trial.Trial,
    trial_dir: Path,
    params: dict[str, Any],
    max_epochs: int,
) -> None:
    lines = [
        "",
        "-" * 80,
        f"OPTUNA TRIAL START | Strategy {strategy} | {percent_label(bundle.percent)}% | Trial {trial.number:03d}",
        "-" * 80,
        f"Study name           : {study_name}",
        f"Run dir              : {trial_dir}",
        f"Max epochs           : {max_epochs}",
        f"Objective metric     : {STUDY_OBJECTIVE_METRIC}",
    ]
    for key in ("head_lr", "encoder_lr", "weight_decay", "dropout_p", "tmax", "entropy_lr"):
        if key in params:
            lines.append(f"{key:22s}: {_format_hparam_value(key, params[key])}")
    tqdm.write("\n".join(lines))

def log_optuna_trial_result(
    *,
    strategy: int,
    bundle: DataBundle,
    trial: optuna.trial.Trial,
    metric_value: float,
    aggregate: dict[str, dict[str, float]],
) -> None:
    tqdm.write(
        f"[Trial {trial.number:03d} | Strategy {strategy} | {percent_label(bundle.percent)}%] "
        f"completed: best_val_iou={metric_value:.4f}, best_test_iou={aggregate['iou']['mean']:.4f}"
    )

def study_paths_for(
    strategy: int,
    percent: float,
    model_config: RuntimeModelConfig | None = None,
) -> tuple[Path, Path, Path]:
    pct_root = RUNS_ROOT / MODEL_NAME / f"pct_{percent_label(percent)}"
    strategy_root = pct_root / strategy_dir_name(strategy, model_config)
    study_root = strategy_root / "study"
    trials_root = strategy_root / "trials"
    return strategy_root, study_root, trials_root

def manual_hparams_key(strategy: int, percent: float) -> str:
    return f"{strategy}:{percent_label(percent)}"

def reset_study_artifacts(strategy: int, percent: float, *, model_config: RuntimeModelConfig) -> None:
    strategy_root, study_root, trials_root = study_paths_for(strategy, percent, model_config)
    removed_any = False
    for path in (study_root, trials_root):
        if path.exists():
            shutil.rmtree(path)
            removed_any = True
    if removed_any:
        print(
            f"[Optuna Reset] Removed cached study artifacts for strategy={strategy}, "
            f"percent={percent_label(percent)}% under {strategy_root}."
        )
    else:
        print(
            f"[Optuna Reset] No existing study artifacts found for strategy={strategy}, "
            f"percent={percent_label(percent)}%."
        )

def pruner_for_run() -> optuna.pruners.BasePruner:
    if USE_TRIAL_PRUNING:
        return optuna.pruners.MedianPruner(n_warmup_steps=TRIAL_PRUNER_WARMUP_STEPS)
    return optuna.pruners.NopPruner()

def run_single_job(
    *,
    strategy: int,
    model_config: RuntimeModelConfig,
    bundle: DataBundle,
    run_dir: Path,
    params: dict[str, Any],
    max_epochs: int,
    trial: optuna.trial.Trial | None,
    strategy2_checkpoint_path: str | Path | None = None,
    resume_checkpoint_path: Path | None = None,
    retrying_from_trial_number: int | None = None,
) -> tuple[dict[str, Any], list[dict[str, Any]], dict[str, dict[str, float]]]:
    set_current_job_params(params)
    DEFAULT_ENTROPY_TARGET_RATIO = float(_job_param("entropy_target_ratio", 0.35))
    DEFAULT_ENTROPY_ALPHA_INIT = float(_job_param("entropy_alpha_init", 0.12))
    ensure_dir(run_dir)
    run_type = "trial" if trial is not None else "final"
    config = {
        "project_dir": str(PROJECT_DIR),
        "data_root": str(DATA_ROOT),
        "run_type": run_type,
        "strategy": strategy,
        "dataset_percent": bundle.percent,
        "dataset_name": bundle.split_payload["dataset_name"],
        "dataset_split_policy": bundle.split_payload["dataset_split_policy"],
        "dataset_splits_path": bundle.split_payload["dataset_splits_path"],
        "split_type": bundle.split_payload["split_type"],
        "train_subset_key": bundle.split_payload["train_subset_key"],
        "normalization_cache_path": bundle.split_payload["normalization_cache_path"],
        "head_lr": params["head_lr"],
        "encoder_lr": params["encoder_lr"],
        "weight_decay": params["weight_decay"],
        "dropout_p": params["dropout_p"],
        "tmax": params["tmax"],
        "entropy_lr": params["entropy_lr"],
        "max_epochs": max_epochs,
        "gamma": DEFAULT_GAMMA,
        "grad_clip_norm": DEFAULT_GRAD_CLIP_NORM,
        "scheduler_factor": SCHEDULER_FACTOR,
        "scheduler_patience": SCHEDULER_PATIENCE,
        "scheduler_threshold": SCHEDULER_THRESHOLD,
        "scheduler_min_lr": SCHEDULER_MIN_LR,
        "execution_mode": EXECUTION_MODE,
        "evaluation_checkpoint_mode": EVAL_CHECKPOINT_MODE,
        "strategy2_checkpoint_mode": STRATEGY2_CHECKPOINT_MODE,
        "train_resume_mode": TRAIN_RESUME_MODE,
        "train_resume_specific_checkpoint": TRAIN_RESUME_SPECIFIC_CHECKPOINT,
    }
    config.update({key: value for key, value in params.items() if key not in config})
    config.update(model_config.to_payload())
    if strategy2_checkpoint_path is not None:
        config["strategy2_checkpoint_path"] = str(Path(strategy2_checkpoint_path))
    if resume_checkpoint_path is not None:
        config["resume_checkpoint_path"] = str(Path(resume_checkpoint_path))
    if retrying_from_trial_number is not None:
        config["retrying_from_trial_number"] = int(retrying_from_trial_number)
    save_json(run_dir / "run_config.json", config)

    model: nn.Module | None = None
    try:
        model, description, _compiled = build_model(
            strategy,
            params["dropout_p"],
            model_config=model_config,
            strategy2_checkpoint_path=strategy2_checkpoint_path,
        )
        summary, history = train_model(
            run_type=run_type,
            model_config=model_config,
            run_config=config,
            model=model,
            description=description,
            strategy=strategy,
            run_dir=run_dir,
            bundle=bundle,
            max_epochs=max_epochs,
            head_lr=params["head_lr"],
            encoder_lr=params["encoder_lr"],
            weight_decay=params["weight_decay"],
            tmax=params["tmax"],
            entropy_lr=params["entropy_lr"],
            entropy_alpha_init=DEFAULT_ENTROPY_ALPHA_INIT,
            entropy_target_ratio=DEFAULT_ENTROPY_TARGET_RATIO,
            critic_loss_weight=DEFAULT_CRITIC_LOSS_WEIGHT,
            ce_weight=DEFAULT_CE_WEIGHT,
            dice_weight=DEFAULT_DICE_WEIGHT,
            dropout_p=params["dropout_p"],
            resume_checkpoint_path=resume_checkpoint_path,
            trial=trial,
        )

        if trial is not None:
            save_json(
                run_dir / "summary.json",
                {
                    "params": params,
                    "best_iou": float(summary["best_val_iou"]),
                    "resumed": bool(resume_checkpoint_path is not None),
                    "retrying_from_trial_number": retrying_from_trial_number,
                },
            )
            return summary, history, {}

        del model
        model = None
        run_cuda_cleanup()

        aggregate, _per_sample = run_evaluation_for_run(
            strategy=strategy,
            percent=bundle.percent,
            bundle=bundle,
            run_dir=run_dir,
            strategy2_checkpoint_path=strategy2_checkpoint_path,
        )
        return summary, history, aggregate
    finally:
        if model is not None:
            del model
            model = None
        run_cuda_cleanup()

def run_study(
    strategy: int,
    bundle: DataBundle,
    *,
    model_config: RuntimeModelConfig,
    strategy2_checkpoint_path: str | Path | None = None,
) -> dict[str, Any]:
    strategy_root, study_root, trials_root = study_paths_for(strategy, bundle.percent, model_config)
    ensure_dir(strategy_root.parent)
    strategy_root = ensure_dir(strategy_root)
    study_root = ensure_dir(strategy_root / "study")
    trials_root = ensure_dir(strategy_root / "trials")
    storage_path = study_root / "study.sqlite3"
    storage = RDBStorage(
        url=f"sqlite:///{storage_path.resolve()}",
        heartbeat_interval=OPTUNA_HEARTBEAT_INTERVAL,
        grace_period=OPTUNA_HEARTBEAT_GRACE_PERIOD,
        failed_trial_callback=RetryFailedTrialCallback(max_retry=OPTUNA_FAILED_TRIAL_MAX_RETRY),
    )
    sampler = optuna.samplers.TPESampler(seed=SEED)
    study_name = (
        f"{MODEL_NAME}_{model_config.backbone_tag()}_pct_{percent_label(bundle.percent)}_strategy_{strategy}"
    )
    study = optuna.create_study(
        study_name=study_name,
        direction=STUDY_DIRECTION,
        sampler=sampler,
        pruner=pruner_for_run(),
        storage=storage,
        load_if_exists=LOAD_EXISTING_STUDIES,
    )
    finished_trials = [trial for trial in study.trials if trial.state.is_finished()]
    completed_trials = [trial for trial in study.trials if trial.state == optuna.trial.TrialState.COMPLETE]
    remaining_trials = max(int(NUM_TRIALS) - len(finished_trials), 0)

    if finished_trials:
        print(
            f"[Optuna Study] Loaded existing study '{study_name}' with "
            f"{len(finished_trials)} finished trial(s), {len(completed_trials)} completed trial(s), "
            f"target={NUM_TRIALS}, remaining={remaining_trials}."
        )
    else:
        print(f"[Optuna Study] Starting new study '{study_name}' with target={NUM_TRIALS} trial(s).")

    def objective(trial: optuna.trial.Trial) -> float:
        trial_dir = ensure_dir(trials_root / f"trial_{trial.number:03d}")
        retried_trial_fn = getattr(RetryFailedTrialCallback, "retried_trial_number", None)
        retrying_from_trial_number = (
            retried_trial_fn(trial) if callable(retried_trial_fn) else trial.system_attrs.get("failed_trial")
        )
        resume_checkpoint_path: Path | None = None
        if retrying_from_trial_number is not None:
            prior_trial_dir = trials_root / f"trial_{int(retrying_from_trial_number):03d}"
            candidate = prior_trial_dir / "checkpoints" / "latest.pt"
            if candidate.exists():
                resume_checkpoint_path = candidate.resolve()
        params = suggest_hyperparameters(trial, strategy)
        log_optuna_trial_start(
            study_name=study_name,
            strategy=strategy,
            bundle=bundle,
            trial=trial,
            trial_dir=trial_dir,
            params=params,
            max_epochs=strategy_epochs(strategy),
        )
        summary: dict[str, Any] | None = None
        _history: list[dict[str, Any]] | None = None
        aggregate: dict[str, dict[str, float]] | None = None
        completed_successfully = False
        run_cuda_cleanup()
        try:
            summary, _history, aggregate = run_single_job(
                strategy=strategy,
                model_config=model_config,
                bundle=bundle,
                run_dir=trial_dir,
                params=params,
                max_epochs=strategy_epochs(strategy),
                trial=trial,
                strategy2_checkpoint_path=strategy2_checkpoint_path,
                resume_checkpoint_path=resume_checkpoint_path,
                retrying_from_trial_number=int(retrying_from_trial_number) if retrying_from_trial_number is not None else None,
            )
            metric_value = float(summary["best_val_iou"])
            completed_successfully = True
            tqdm.write(
                f"[Trial {trial.number:03d} | Strategy {strategy} | {percent_label(bundle.percent)}%] "
                f"completed: best_val_iou={metric_value:.4f}"
            )
            return metric_value
        finally:
            if aggregate is not None:
                del aggregate
                aggregate = None
            if _history is not None:
                del _history
                _history = None
            if summary is not None:
                del summary
                summary = None
            if completed_successfully:
                prune_optuna_trial_dir(trial_dir)
            run_cuda_cleanup(context=f"trial {trial.number:03d} boundary")

    if remaining_trials > 0:
        study.optimize(objective, n_trials=remaining_trials, show_progress_bar=True)
    else:
        print(
            f"[Optuna Study] Skipping optimization for '{study_name}' because the study already reached "
            f"the configured cap of {NUM_TRIALS} finished trial(s)."
        )

    completed_trials = [trial for trial in study.trials if trial.state == optuna.trial.TrialState.COMPLETE]
    if not completed_trials:
        raise RuntimeError(
            f"Study '{study_name}' has no completed trials, so best params cannot be resolved. "
            f"Finished trials={len([trial for trial in study.trials if trial.state.is_finished()])}, "
            f"configured cap={NUM_TRIALS}."
        )

    best_params = dict(study.best_trial.params)
    if strategy == 2:
        best_params.setdefault("tmax", DEFAULT_TMAX)
        best_params.setdefault("entropy_lr", DEFAULT_ENTROPY_LR)
    save_json(study_root / "best_params.json", best_params)
    save_json(
        study_root / "summary.json",
        {
            "best_params": best_params,
            "best_iou": float(study.best_value),
            "finished_trials": len([trial for trial in study.trials if trial.state.is_finished()]),
            "completed_trials": len(completed_trials),
            "target_trials": int(NUM_TRIALS),
            "remaining_trials": max(int(NUM_TRIALS) - len([trial for trial in study.trials if trial.state.is_finished()]), 0),
        },
    )
    prune_optuna_study_dir(study_root)
    return best_params

def run_final_training(
    strategy: int,
    bundle: DataBundle,
    params: dict[str, Any],
    *,
    model_config: RuntimeModelConfig,
    strategy2_checkpoint_path: str | Path | None = None,
) -> None:
    final_root = final_root_for_strategy(strategy, bundle.percent, model_config)
    if SKIP_EXISTING_FINALS and (final_root / "summary.json").exists():
        print(f"Skipping existing final run: {final_root}")
        return
    save_json(final_root / "best_params.json", params)
    resume_checkpoint_path = resolve_train_resume_checkpoint_path(final_root)
    run_single_job(
        strategy=strategy,
        model_config=model_config,
        bundle=bundle,
        run_dir=final_root,
        params=params,
        max_epochs=strategy_epochs(strategy),
        trial=None,
        strategy2_checkpoint_path=strategy2_checkpoint_path,
        resume_checkpoint_path=resume_checkpoint_path,
    )

def print_environment_summary(model_config: RuntimeModelConfig) -> None:
    banner("RUNTIME SUMMARY")
    images_dir, annotations_dir = current_dataset_dirs()
    print(f"Project dir           : {PROJECT_DIR}")
    print(f"Data root             : {DATA_ROOT}")
    print(f"Runs root             : {RUNS_ROOT}")
    print(f"Dataset name          : {current_dataset_name()}")
    if current_dataset_name() == "BUSI_with_classes":
        print(f"Dataset split policy  : {current_busi_with_classes_split_policy()}")
    print(f"Images dir            : {images_dir}")
    print(f"Masks dir             : {annotations_dir}")
    print(f"Dataset splits json   : {current_dataset_splits_json_path()}")
    print(f"Split type            : {SPLIT_TYPE}")
    print(f"Device                : {DEVICE}")
    print(f"Device source         : {DEVICE_FALLBACK_SOURCE}")
    print(f"Model name            : {MODEL_NAME}")
    print(f"Seed                  : {SEED}")
    print(f"PyTorch version       : {torch.__version__}")
    print(f"Batch size            : {BATCH_SIZE}")
    print(f"Use AMP               : {USE_AMP}")
    print(f"Num workers           : {NUM_WORKERS}")
    print(f"Pin memory            : {USE_PIN_MEMORY}")
    print(f"CuDNN deterministic   : {torch.backends.cudnn.deterministic}")
    print(f"CuDNN benchmark       : {torch.backends.cudnn.benchmark}")

    if DEVICE.type == "cuda":
        props = torch.cuda.get_device_properties(0)
        print(f"GPU                   : {torch.cuda.get_device_name(0)}")
    print(f"GPU VRAM              : {props.total_memory / (1024 ** 3):.2f} GB")
    print(f"AMP dtype             : {resolve_amp_dtype(AMP_DTYPE)}")
    print(f"Trial pruning         : {USE_TRIAL_PRUNING}")
    print(f"Backbone family       : {model_config.backbone_family}")
    if model_config.backbone_family == "custom_vgg":
        print(f"VGG feature scales    : {model_config.vgg_feature_scales}")
        print(f"VGG feature dilation  : {model_config.vgg_feature_dilation}")
    else:
        print(f"SMP encoder           : {model_config.smp_encoder_name}")
        print(f"SMP encoder depth     : {model_config.smp_encoder_depth}")
        print(f"SMP encoder proj dim  : {model_config.smp_encoder_proj_dim}")
        print(f"SMP decoder           : {model_config.smp_decoder_type}")
    print(f"Strategies            : {STRATEGIES}")
    print(f"Dataset percents      : {DATASET_PERCENTS}")
    print_imagenet_normalization_status()
    print(f"Trials per study      : {NUM_TRIALS}")
    print(f"Execution mode        : {EXECUTION_MODE}")
    print(f"Run Optuna            : {RUN_OPTUNA}")
    print(f"Use saved best params : {USE_EXISTING_BEST_PARAMS_WHEN_OPTUNA_OFF}")
    print(f"Reset studies/run     : {RESET_ALL_STUDIES_EACH_RUN}")
    print(f"Load existing studies : {LOAD_EXISTING_STUDIES}")
    print(f"Eval ckpt selector    : {EVAL_CHECKPOINT_MODE}")
    print(f"S2 ckpt selector      : {STRATEGY2_CHECKPOINT_MODE}")
    print(f"S3 bootstrap from S2  : {STRATEGY3_BOOTSTRAP_FROM_STRATEGY2}")
    print(f"Train resume mode     : {TRAIN_RESUME_MODE}")
    print(f"Verbose epoch log     : {VERBOSE_EPOCH_LOG}")
    print(f"Validate every epochs : {VALIDATE_EVERY_N_EPOCHS}")
    print(f"Smoke test enabled    : {RUN_SMOKE_TEST}")
    print(f"Overfit test enabled  : {RUN_OVERFIT_TEST}")
    print(f"Overfit batches       : {OVERFIT_N_BATCHES}")
    print(f"Overfit epochs        : {OVERFIT_N_EPOCHS}")

def maybe_run_strategy_smoke_test(
    *,
    strategy: int,
    model_config: RuntimeModelConfig,
    bundle: DataBundle,
    strategy2_checkpoint_path: str | Path | None = None,
) -> None:
    if not RUN_SMOKE_TEST:
        return
    if EXECUTION_MODE == "eval_only":
        return
    run_smoke_test(
        strategy=strategy,
        model_config=model_config,
        bundle=bundle,
        smoke_root=strategy_root_for_percent(strategy, bundle.percent, model_config) / "smoke_test",
        strategy2_checkpoint_path=strategy2_checkpoint_path,
    )

def main() -> int:
    banner("MLR ALL STRATEGIES BAYES RUNNER")
    set_global_seed(SEED)
    model_config = current_model_config()
    dataset_name = current_dataset_name()
    images_dir, annotations_dir = current_dataset_dirs()
    if EXECUTION_MODE not in {"train_eval", "eval_only"}:
        raise ValueError(f"EXECUTION_MODE must be 'train_eval' or 'eval_only', got {EXECUTION_MODE!r}")
    if SPLIT_TYPE not in SUPPORTED_SPLIT_TYPES:
        raise ValueError(f"SPLIT_TYPE must be one of {SUPPORTED_SPLIT_TYPES}, got {SPLIT_TYPE!r}")
    if not images_dir.exists() or not annotations_dir.exists():
        raise FileNotFoundError(
            f"Expected dataset directories for {dataset_name} under {DATA_ROOT}: "
            f"images={images_dir}, masks={annotations_dir}"
        )
    ensure_specific_checkpoint_scope("EVAL_CHECKPOINT_MODE", EVAL_CHECKPOINT_MODE)
    ensure_specific_checkpoint_scope("STRATEGY2_CHECKPOINT_MODE", STRATEGY2_CHECKPOINT_MODE)
    ensure_specific_checkpoint_scope("TRAIN_RESUME_MODE", TRAIN_RESUME_MODE)

    print_environment_summary(model_config)
    split_registry, split_source = load_or_create_dataset_splits(
        images_dir=images_dir,
        annotations_dir=annotations_dir,
        split_json_path=current_dataset_splits_json_path(),
        train_fractions=DATASET_PERCENTS,
        seed=SEED,
    )

    bundles: dict[float, DataBundle] = {}
    for percent in DATASET_PERCENTS:
        bundles[percent] = build_data_bundle(percent, split_registry, split_source)

    if RUN_OVERFIT_TEST:
        run_configured_overfit_tests(bundles, model_config=model_config)
        banner("OVERFIT TESTS COMPLETE")
        return 0

    if RESET_ALL_STUDIES_EACH_RUN:
        if RUN_OPTUNA:
            banner("RESETTING OPTUNA STUDIES")
            for strategy in STRATEGIES:
                for percent in DATASET_PERCENTS:
                    reset_study_artifacts(strategy, percent, model_config=model_config)
        else:
            print("[Optuna Reset] Skipped because RUN_OPTUNA=False.")

    try:
        for percent in DATASET_PERCENTS:
            banner(f"PERCENT STAGE | {percent_label(percent)}%")
            bundle = bundles[percent]
            for strategy in STRATEGIES:
                strategy2_checkpoint_path = None
                if strategy in (4, 5) or (
                    EXECUTION_MODE == "train_eval" and strategy == 3 and STRATEGY3_BOOTSTRAP_FROM_STRATEGY2
                ):
                    strategy2_checkpoint_path = resolve_strategy2_checkpoint_path(strategy, percent, model_config)

                if EXECUTION_MODE == "train_eval":
                    maybe_run_strategy_smoke_test(
                        strategy=strategy,
                        model_config=model_config,
                        bundle=bundle,
                        strategy2_checkpoint_path=strategy2_checkpoint_path,
                    )
                    params = resolve_job_params(
                        strategy,
                        percent,
                        bundle,
                        model_config=model_config,
                        strategy2_checkpoint_path=strategy2_checkpoint_path,
                    )
                    banner(f"FINAL RETRAIN | Strategy {strategy} | {percent_label(percent)}%")
                    run_final_training(
                        strategy,
                        bundle,
                        params,
                        model_config=model_config,
                        strategy2_checkpoint_path=strategy2_checkpoint_path,
                    )
                else:
                    banner(f"EVAL ONLY | Strategy {strategy} | {percent_label(percent)}%")
                    run_evaluation_for_run(
                        strategy=strategy,
                        percent=percent,
                        bundle=bundle,
                        run_dir=final_root_for_strategy(strategy, percent, model_config),
                        strategy2_checkpoint_path=strategy2_checkpoint_path,
                    )
    except Exception:
        banner("RUN FAILED")
        traceback.print_exc()
        return 1

    banner("ALL DONE")
    return 0

if __name__ == "__main__":
    main()