"""Load the IWM inference-only safetensors release; never loads private data.

Use with the source checkout pinned in manifest.json. No training is dispatched.
"""
import hashlib
import json
from pathlib import Path
import sys

import torch
from safetensors.torch import load_file

SOURCE_COMMIT = "516a71c252ff8f895dcd036bb343c996254c7a76"
BANK = ("feature_mean", "feature_std", "coefficient", "action_mean",
        "expert_weight", "expert_mean", "residual_std")
GUIDE = ("probe_mean", "probe_std", "probe_weight", "probe_bias",
         "class_centers", "within_std", "relation_keys", "relation_labels", "relation_ids")
EXCLUDED = ("prototypes", "amplitude_limit", "variance") + tuple("guide_" + k for k in GUIDE)


def digest(path):
    h = hashlib.sha256()
    with Path(path).open("rb") as f:
        for block in iter(lambda: f.read(8 * 1024**2), b""):
            h.update(block)
    return h.hexdigest()


def source_paths(source_root):
    root = Path(source_root).resolve()
    archive = root / "experiments/videorae3d"
    paths = [archive / "cg2_dualcontext", archive / "cg2_ng",
             archive / "eeg", archive / "teacher", archive, root / "src", root / "scripts"]
    if not all(p.is_dir() for p in paths):
        raise ValueError("Use the complete pinned intention-world-model source checkout")
    for path in reversed(paths):
        if str(path) not in sys.path:
            sys.path.insert(0, str(path))
    return archive


def prepared_from_state(state, prefix="", include_training=False):
    """Constructor adapter, not a replacement for private training preparation."""
    normal = {"mean": state[prefix + "normal_mean"], "std": state[prefix + "normal_std"]}
    bank = {k: state[prefix + k] for k in BANK}
    if include_training:
        bank.update({k: state[prefix + k] for k in ("prototypes", "amplitude_limit", "variance")})
        guide = {k: state[prefix + "guide_" + k] for k in GUIDE}
    else:
        # The archived constructor registers training-only buffers. These empty
        # placeholders are deleted before strict loading and never used in forward.
        guide = dict(probe_mean=torch.zeros(256), probe_std=torch.ones(256),
                     probe_weight=torch.zeros(5, 256), probe_bias=torch.zeros(5),
                     class_centers=torch.zeros(5, 256), within_std=torch.ones(256),
                     relation_keys=torch.empty(0, 256),
                     relation_labels=torch.empty(0, dtype=torch.long),
                     relation_ids=torch.empty(0, dtype=torch.long))
    return dict(normalizers={"target": normal}, bank=bank, classguide=guide)


def build_model(config, state, source_root, original=False):
    # safetensors stores contiguous values. Preserve archived dense buffer layout
    # too: a transposed ridge matrix otherwise selects a different BLAS path.
    state = dict(state)
    for name, strides in config.get("tensor_layouts", {}).items():
        t = state[name]
        if len(strides) != t.ndim or any(type(x) is not int or x <= 0 for x in strides):
            raise ValueError("Invalid tensor layout")
        axes = sorted(zip(t.shape, strides), key=lambda pair: pair[1])
        span = 1
        for size, stride in axes:
            if size > 1:
                if stride != span:
                    raise ValueError("Only dense, non-overlapping layouts are supported")
                span *= size
        if tuple(t.stride()) != tuple(strides):
            state[name] = torch.empty_strided(t.shape, strides, dtype=t.dtype).copy_(t)
    source_paths(source_root)
    kind = config["architecture"]
    if kind == "Teacher3D":
        from model import Teacher3D
        m = Teacher3D()
    else:
        if kind not in ("CG2-NG", "NG-D"):
            raise ValueError("Unsupported released architecture")
        cfg = dict(config["model_config"])
        if not (cfg["timing"] == "target" and cfg["method"] == "CG2"
                and cfg["variant"] == "CG2-NG" and cfg["representation"] == "tokens"
                and cfg["head"] == "continuous"):
            raise ValueError("Unexpected model protocol")
        prefix = "base." if kind == "NG-D" else ""
        prepared = prepared_from_state(state, prefix, include_training=original)
        if kind == "NG-D":
            assert not cfg["past_aux"] and cfg["separate_motion"]
            from dual_models import Controller
        else:
            from noise_models import Controller
        m = Controller(cfg, prepared)
        if not original:
            base = m.base if kind == "NG-D" else m
            for name in EXCLUDED:
                if name in base._buffers:
                    delattr(base, name)
    m.load_state_dict(state, strict=True)
    return m.eval().requires_grad_(False)


def load_checkpoint(release_root, model_id, source_root, device="cpu", verify=True):
    """Return an eval-only Teacher3D or EEG controller.

    Controllers take (z_current[B,1024,64], eeg_raw[B,32,500]) and return
    (class_logits[B,5], action[B,1,256]). Zero is raw zero, before normalization.
    Use teacher.fdm(z_current, action) for the predicted H1 latent.
    Training losses/probes and optimizer-resume state are intentionally absent.
    """
    root = Path(release_root).resolve()
    manifest = json.loads((root / "manifest.json").read_text())
    entry = next((r for r in manifest["models"] if r["id"] == model_id), None)
    if entry is None:
        raise ValueError("Unknown model_id; see manifest.json")
    weights, config = root / entry["weights"], root / entry["config"]
    if verify:
        for p in (weights, config):
            rel = p.relative_to(root).as_posix()
            if digest(p) != manifest["files"][rel]["sha256"]:
                raise ValueError("Release checksum mismatch: " + rel)
    cfg = json.loads(config.read_text())
    state = load_file(str(weights), device="cpu")
    return build_model(cfg, state, source_root).to(device)