scripts/research_direction_extension_tasks.py

#!/usr/bin/env python3
"""Run one extra data-backed probe for each Ropedia research direction.

These tasks reuse the committed single-episode feature tensor generated by
`episode_task_suite.py`. They are extension probes, not claims that the full
research directions are solved.
"""

from __future__ import annotations

import argparse
import csv
import html
import json
import math
from collections import OrderedDict
from pathlib import Path
from typing import Any

import numpy as np


ROOT = Path(__file__).resolve().parents[1]
RESULTS = ROOT / "results" / "episode_task_suite"
OUT_DIR = RESULTS / "research_direction_extensions"
DOCS_DATA = ROOT / "docs" / "data"
CHARTS = ROOT / "docs" / "assets" / "charts"

WINDOWS_NPZ = RESULTS / "shared_windows.npz"
WINDOWS_CSV = RESULTS / "windows.csv"
FEATURE_MANIFEST = RESULTS / "feature_manifest.json"


TASK_SPECS: OrderedDict[str, dict[str, Any]] = OrderedDict(
    [
        (
            "body_motion_intensity",
            {
                "direction": "A",
                "direction_name": "Human Modeling & Motion Understanding",
                "name": "Body/hand motion intensity",
                "family": "classification",
                "case_study": "A window with a fast reach or pour should be classified as high motion; a steady holding window should be low motion.",
                "input": "Current non-mocap feature blocks: video, depth, camera pose/rotation, IMU, SLAM, calibration, and language context.",
                "middle_process": "Compute the target from hand/body joint changes between neighboring windows, hide the mocap blocks from the input, then classify high versus low motion using the train-set median as the threshold.",
                "output": "Binary label: high_motion or low_motion.",
                "minimal_baseline": "Ridge classifier on standardized non-mocap features.",
                "neural_baseline": "One-hidden-layer MLP binary classifier on the same input features.",
                "metric_name": "macro-F1",
                "metric_key": "macro_f1",
                "metric_direction": "higher",
                "current_limit": "This is a motion-energy proxy, not a SMPL/MANO body model or a generative motion prior.",
            },
        ),
        (
            "multi_view_consistency_retrieval",
            {
                "direction": "B",
                "direction_name": "3D/4D Reconstruction & Neural Rendering",
                "name": "Multi-view consistency retrieval",
                "family": "retrieval",
                "case_study": "Given the fisheye camera features for a pouring moment, retrieve the synchronized stereo-left view from the same time window.",
                "input": "Query side: fisheye_cam0 video feature block. Candidate side: stereo_left video feature block from held-out windows.",
                "middle_process": "Learn a projection from one camera-view feature space into another, then rank held-out candidate windows by cosine similarity.",
                "output": "Ranked candidate windows; the correct synchronized view should rank near the top.",
                "minimal_baseline": "Ridge projection followed by cosine nearest-neighbor retrieval.",
                "neural_baseline": "One-hidden-layer MLP projection followed by the same cosine retrieval evaluator.",
                "metric_name": "MRR",
                "metric_key": "mrr",
                "metric_direction": "higher",
                "current_limit": "This checks calibrated multi-view signal, but it is still feature retrieval, not NeRF, Gaussian Splatting, or novel-view synthesis.",
            },
        ),
        (
            "action_phase_progress",
            {
                "direction": "C",
                "direction_name": "Egocentric Vision & Interaction",
                "name": "Action phase progress",
                "family": "regression",
                "case_study": "Inside a Pour coffee action segment, estimate whether the current window is near the beginning, middle, or end of that action.",
                "input": "Current non-caption multimodal feature vector, so the label text cannot be copied directly from the language block.",
                "middle_process": "Convert contiguous action-label runs into a normalized 0-to-1 progress target, train on earlier windows, and regress progress for later windows.",
                "output": "A scalar progress value between 0.0 and 1.0 for the current action segment.",
                "minimal_baseline": "Ridge regressor on standardized non-caption features.",
                "neural_baseline": "One-hidden-layer MLP regressor on the same input features.",
                "metric_name": "MAE",
                "metric_key": "mae",
                "metric_direction": "lower",
                "current_limit": "This is an action-structure probe inside one episode, not a general intent model across homes, people, or tasks.",
            },
        ),
        (
            "ego_motion_forecast",
            {
                "direction": "D",
                "direction_name": "Scene Reconstruction & World Modeling",
                "name": "Short-horizon ego-motion forecast",
                "family": "forecast",
                "case_study": "From the current sensors, predict how the camera translation will change over the next 20 frames while the wearer moves through the scene.",
                "input": "Current multimodal features excluding the camera-translation block and caption text.",
                "middle_process": "Build a future target from camera-translation difference at a four-window horizon, then regress that future ego-motion delta from current sensors.",
                "output": "A future camera-translation delta vector.",
                "minimal_baseline": "Ridge regressor with a 20-frame forecast horizon.",
                "neural_baseline": "One-hidden-layer MLP regressor with the same horizon and split.",
                "metric_name": "MAE",
                "metric_key": "mae",
                "metric_direction": "lower",
                "current_limit": "This is a compact world-model proxy; it does not build a persistent map, scene graph, or object permanence model.",
            },
        ),
    ]
)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Run four research-direction extension probes.")
    parser.add_argument("--results-dir", type=Path, default=RESULTS)
    parser.add_argument("--output-dir", type=Path, default=OUT_DIR)
    parser.add_argument("--train-fraction", type=float, default=0.70)
    parser.add_argument("--ridge-l2", type=float, default=10.0)
    parser.add_argument("--seed", type=int, default=7)
    parser.add_argument("--future-windows", type=int, default=4)
    parser.add_argument("--neural-epochs", type=int, default=25)
    parser.add_argument("--neural-hidden-dim", type=int, default=128)
    parser.add_argument("--neural-batch-size", type=int, default=128)
    parser.add_argument("--neural-learning-rate", type=float, default=1e-3)
    parser.add_argument("--neural-weight-decay", type=float, default=1e-4)
    parser.add_argument("--skip-neural", action="store_true")
    return parser.parse_args()


def write_json(path: Path, payload: dict[str, Any] | list[Any]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(payload, indent=2), encoding="utf-8")


def write_csv(path: Path, rows: list[dict[str, Any]], fieldnames: list[str]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", newline="", encoding="utf-8") as handle:
        writer = csv.DictWriter(handle, fieldnames=fieldnames, lineterminator="\n")
        writer.writeheader()
        writer.writerows(rows)


def load_windows_csv(path: Path) -> list[dict[str, str]]:
    with path.open("r", newline="", encoding="utf-8") as handle:
        return list(csv.DictReader(handle))


def load_inputs(results_dir: Path) -> tuple[np.ndarray, np.ndarray, np.ndarray, list[dict[str, str]], list[dict[str, Any]]]:
    npz_path = results_dir / "shared_windows.npz"
    windows_path = results_dir / "windows.csv"
    manifest_path = results_dir / "feature_manifest.json"
    if not npz_path.exists():
        raise FileNotFoundError(f"Missing {npz_path}. Run scripts/episode_task_suite.py first.")
    z = np.load(npz_path, allow_pickle=False)
    X = np.asarray(z["X"], dtype=np.float32)
    starts = np.asarray(z["starts"], dtype=np.int64)
    ends = np.asarray(z["ends"], dtype=np.int64)
    X = np.nan_to_num(X, nan=0.0, posinf=0.0, neginf=0.0)
    rows = load_windows_csv(windows_path)
    manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
    if len(rows) != len(X):
        raise ValueError(f"windows.csv has {len(rows)} rows but shared_windows.npz has {len(X)} windows.")
    return X, starts, ends, rows, manifest


def block_indices(manifest: list[dict[str, Any]], include: list[str] | None = None, exclude: list[str] | None = None) -> np.ndarray:
    include = include or []
    exclude = exclude or []
    idxs: list[int] = []
    for block in manifest:
        name = str(block["name"])
        if include and not any(name == p or name.startswith(p) for p in include):
            continue
        if exclude and any(name == p or name.startswith(p) for p in exclude):
            continue
        idxs.extend(range(int(block["start"]), int(block["end"])))
    return np.asarray(idxs, dtype=np.int64)


def chronological_split(n: int, train_fraction: float) -> tuple[np.ndarray, np.ndarray]:
    if n < 2:
        raise ValueError("Need at least two examples.")
    split = int(round(n * train_fraction))
    split = max(1, min(split, n - 1))
    return np.arange(split, dtype=np.int64), np.arange(split, n, dtype=np.int64)


def standardize_train_test(X_train: np.ndarray, X_test: np.ndarray) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
    mean = X_train.mean(axis=0, dtype=np.float64).astype(np.float32)
    std = X_train.std(axis=0, dtype=np.float64).astype(np.float32)
    std[std < 1e-6] = 1.0
    return (X_train - mean) / std, (X_test - mean) / std, mean, std


def ridge_predict(
    X_train: np.ndarray,
    Y_train: np.ndarray,
    X_test: np.ndarray,
    *,
    l2: float,
    standardize_y: bool,
) -> np.ndarray:
    Xtr, Xte, _, _ = standardize_train_test(X_train.astype(np.float32), X_test.astype(np.float32))
    Y = np.asarray(Y_train, dtype=np.float32)
    if Y.ndim == 1:
        Y = Y[:, None]
    if standardize_y:
        y_mean = Y.mean(axis=0, dtype=np.float64).astype(np.float32)
        y_std = Y.std(axis=0, dtype=np.float64).astype(np.float32)
        y_std[y_std < 1e-6] = 1.0
        Y_work = (Y - y_mean) / y_std
    else:
        y_mean = np.zeros(Y.shape[1], dtype=np.float32)
        y_std = np.ones(Y.shape[1], dtype=np.float32)
        Y_work = Y

    K = Xtr @ Xtr.T
    K.flat[:: K.shape[0] + 1] += float(l2)
    alpha = np.linalg.solve(K.astype(np.float64), Y_work.astype(np.float64)).astype(np.float32)
    pred = (Xte @ Xtr.T @ alpha).astype(np.float32)
    return pred * y_std + y_mean


def ridge_classifier(X_train: np.ndarray, y_train: np.ndarray, X_test: np.ndarray, l2: float) -> tuple[np.ndarray, np.ndarray]:
    classes = np.asarray(sorted(set(int(v) for v in y_train)), dtype=np.int64)
    class_to_col = {int(cls): i for i, cls in enumerate(classes)}
    Y = np.zeros((len(y_train), len(classes)), dtype=np.float32)
    for row, label in enumerate(y_train):
        Y[row, class_to_col[int(label)]] = 1.0
    scores = ridge_predict(X_train, Y, X_test, l2=l2, standardize_y=False)
    pred = classes[np.argmax(scores, axis=1)]
    return pred.astype(np.int64), scores


def binary_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict[str, float | int]:
    y_true = y_true.astype(np.int64)
    y_pred = y_pred.astype(np.int64)
    accuracy = float(np.mean(y_true == y_pred))
    per_class_f1 = []
    for cls in (0, 1):
        tp = int(np.sum((y_true == cls) & (y_pred == cls)))
        fp = int(np.sum((y_true != cls) & (y_pred == cls)))
        fn = int(np.sum((y_true == cls) & (y_pred != cls)))
        precision = tp / max(tp + fp, 1)
        recall = tp / max(tp + fn, 1)
        f1 = 2 * precision * recall / max(precision + recall, 1e-12)
        per_class_f1.append(f1)
    return {
        "accuracy": accuracy,
        "macro_f1": float(np.mean(per_class_f1)),
        "positive_rate_true": float(np.mean(y_true)),
        "positive_rate_pred": float(np.mean(y_pred)),
        "num_test": int(len(y_true)),
    }


def regression_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict[str, float | int]:
    y_true = np.asarray(y_true, dtype=np.float32)
    y_pred = np.asarray(y_pred, dtype=np.float32)
    err = y_pred - y_true
    mse = float(np.mean(err * err))
    mae = float(np.mean(np.abs(err)))
    denom = float(np.sum((y_true - y_true.mean(axis=0, keepdims=True)) ** 2))
    numer = float(np.sum(err * err))
    r2 = 1.0 - numer / max(denom, 1e-12)
    return {"mse": mse, "mae": mae, "r2": r2, "num_test": int(len(y_true))}


def row_normalize(X: np.ndarray) -> np.ndarray:
    denom = np.linalg.norm(X, axis=1, keepdims=True)
    denom[denom < 1e-8] = 1.0
    return X / denom


def retrieval_metrics(query_pred: np.ndarray, target_test: np.ndarray) -> tuple[dict[str, float | int], list[dict[str, Any]]]:
    Q = row_normalize(np.asarray(query_pred, dtype=np.float32))
    T = row_normalize(np.asarray(target_test, dtype=np.float32))
    sims = Q @ T.T
    ranks = []
    rows = []
    for i in range(sims.shape[0]):
        order = np.argsort(-sims[i])
        rank = int(np.flatnonzero(order == i)[0]) + 1
        ranks.append(rank)
        rows.append(
            {
                "test_position": i,
                "true_rank": rank,
                "top_candidate_position": int(order[0]),
                "top_candidate_score": float(sims[i, order[0]]),
                "true_score": float(sims[i, i]),
            }
        )
    ranks_array = np.asarray(ranks, dtype=np.float32)
    metrics = {
        "mrr": float(np.mean(1.0 / ranks_array)),
        "top1": float(np.mean(ranks_array <= 1)),
        "top5": float(np.mean(ranks_array <= 5)),
        "top10": float(np.mean(ranks_array <= 10)),
        "median_rank": float(np.median(ranks_array)),
        "num_test": int(len(ranks)),
    }
    return metrics, rows


def choose_score(task: str, metrics: dict[str, Any]) -> float:
    spec = TASK_SPECS[task]
    value = float(metrics[spec["metric_key"]])
    if spec["metric_direction"] == "higher":
        return value
    return max(0.0, 1.0 - value)


def train_neural(
    X_train: np.ndarray,
    Y_train: np.ndarray,
    X_test: np.ndarray,
    *,
    task_type: str,
    args: argparse.Namespace,
) -> tuple[np.ndarray, dict[str, Any]]:
    try:
        import torch
        from torch import nn
        from torch.utils.data import DataLoader, TensorDataset
    except Exception as exc:  # pragma: no cover - depends on optional torch install
        return np.empty((len(X_test), 0), dtype=np.float32), {"available": False, "reason": f"torch unavailable: {exc}"}

    rng = np.random.default_rng(args.seed)
    torch.manual_seed(args.seed)
    Xtr, Xte, _, _ = standardize_train_test(X_train.astype(np.float32), X_test.astype(np.float32))
    Y = np.asarray(Y_train, dtype=np.float32)
    if Y.ndim == 1:
        Y = Y[:, None]
    if task_type == "classification":
        Y_work = Y
        y_mean = np.zeros(Y.shape[1], dtype=np.float32)
        y_std = np.ones(Y.shape[1], dtype=np.float32)
    else:
        y_mean = Y.mean(axis=0, dtype=np.float64).astype(np.float32)
        y_std = Y.std(axis=0, dtype=np.float64).astype(np.float32)
        y_std[y_std < 1e-6] = 1.0
        Y_work = (Y - y_mean) / y_std

    device = torch.device("cpu")
    model = nn.Sequential(
        nn.Linear(Xtr.shape[1], args.neural_hidden_dim),
        nn.GELU(),
        nn.Dropout(0.08),
        nn.Linear(args.neural_hidden_dim, Y_work.shape[1]),
    ).to(device)
    if task_type == "classification":
        loss_fn = nn.BCEWithLogitsLoss()
    else:
        loss_fn = nn.MSELoss()
    opt = torch.optim.AdamW(model.parameters(), lr=args.neural_learning_rate, weight_decay=args.neural_weight_decay)

    order = np.arange(len(Xtr))
    dataset = TensorDataset(torch.from_numpy(Xtr), torch.from_numpy(Y_work.astype(np.float32)))
    loader = DataLoader(dataset, batch_size=args.neural_batch_size, shuffle=False)
    history = []
    for epoch in range(args.neural_epochs):
        rng.shuffle(order)
        if len(order) == len(dataset):
            X_epoch = torch.from_numpy(Xtr[order])
            Y_epoch = torch.from_numpy(Y_work.astype(np.float32)[order])
            loader = DataLoader(TensorDataset(X_epoch, Y_epoch), batch_size=args.neural_batch_size, shuffle=False)
        model.train()
        total_loss = 0.0
        total_seen = 0
        for xb, yb in loader:
            xb = xb.to(device)
            yb = yb.to(device)
            opt.zero_grad(set_to_none=True)
            loss = loss_fn(model(xb), yb)
            loss.backward()
            opt.step()
            total_loss += float(loss.item()) * len(xb)
            total_seen += len(xb)
        history.append(total_loss / max(total_seen, 1))

    model.eval()
    with torch.no_grad():
        raw = model(torch.from_numpy(Xte).to(device)).cpu().numpy().astype(np.float32)
    if task_type == "classification":
        pred = 1.0 / (1.0 + np.exp(-raw))
    else:
        pred = raw * y_std + y_mean
    return pred, {"available": True, "epochs": args.neural_epochs, "hidden_dim": args.neural_hidden_dim, "loss_history": history}


def action_progress_targets(rows: list[dict[str, str]]) -> np.ndarray:
    labels = [row.get("action_label", "") or "" for row in rows]
    progress = np.zeros(len(labels), dtype=np.float32)
    start = 0
    while start < len(labels):
        end = start + 1
        while end < len(labels) and labels[end] == labels[start]:
            end += 1
        length = end - start
        if length > 1:
            progress[start:end] = np.linspace(0.0, 1.0, length, dtype=np.float32)
        start = end
    return progress


def task_body_motion_intensity(X: np.ndarray, rows: list[dict[str, str]], manifest: list[dict[str, Any]], args: argparse.Namespace) -> dict[str, Any]:
    mocap_idx = block_indices(manifest, include=["hand_left_joints", "hand_right_joints", "body_joints"])
    input_idx = block_indices(manifest, exclude=["hand_left_joints", "hand_right_joints", "body_joints", "body_contacts"])
    valid = np.arange(1, len(X), dtype=np.int64)
    motion = np.linalg.norm(X[valid][:, mocap_idx] - X[valid - 1][:, mocap_idx], axis=1)
    train_local, test_local = chronological_split(len(valid), args.train_fraction)
    threshold = float(np.median(motion[train_local]))
    y = (motion >= threshold).astype(np.int64)
    Xv = X[valid][:, input_idx]
    y_pred_min, scores = ridge_classifier(Xv[train_local], y[train_local], Xv[test_local], args.ridge_l2)
    min_metrics = binary_metrics(y[test_local], y_pred_min)
    min_rows = []
    for local_pos, pred, score_pair in zip(test_local, y_pred_min, scores):
        idx = int(valid[int(local_pos)])
        min_rows.append(
            {
                "window_index": idx,
                "center_frame": rows[idx]["center_frame"],
                "motion_energy": float(motion[int(local_pos)]),
                "true_label": "high_motion" if y[int(local_pos)] else "low_motion",
                "pred_label": "high_motion" if int(pred) else "low_motion",
                "score_low": float(score_pair[0]) if len(score_pair) > 0 else "",
                "score_high": float(score_pair[1]) if len(score_pair) > 1 else "",
            }
        )

    neural = {"available": False, "reason": "skipped by flag"}
    neural_metrics = None
    neural_rows: list[dict[str, Any]] = []
    if not args.skip_neural:
        prob, neural = train_neural(Xv[train_local], y[train_local].astype(np.float32), Xv[test_local], task_type="classification", args=args)
        if neural.get("available") and prob.size:
            pred = (prob[:, 0] >= 0.5).astype(np.int64)
            neural_metrics = binary_metrics(y[test_local], pred)
            for local_pos, p, pr in zip(test_local, pred, prob[:, 0]):
                idx = int(valid[int(local_pos)])
                neural_rows.append(
                    {
                        "window_index": idx,
                        "center_frame": rows[idx]["center_frame"],
                        "motion_energy": float(motion[int(local_pos)]),
                        "true_label": "high_motion" if y[int(local_pos)] else "low_motion",
                        "pred_label": "high_motion" if int(p) else "low_motion",
                        "prob_high": float(pr),
                    }
                )

    write_csv(OUT_DIR / "body_motion_intensity_minimal_predictions.csv", min_rows, list(min_rows[0].keys()))
    if neural_rows:
        write_csv(OUT_DIR / "body_motion_intensity_neural_predictions.csv", neural_rows, list(neural_rows[0].keys()))
    return {
        "train_windows": int(len(train_local)),
        "test_windows": int(len(test_local)),
        "target_threshold_train_median": threshold,
        "input_dim": int(len(input_idx)),
        "target_source": "hand/body joint delta between neighboring windows",
        "minimal": min_metrics,
        "neural_mlp": neural_metrics,
        "neural_training": neural,
    }


def task_multi_view_retrieval(X: np.ndarray, rows: list[dict[str, str]], manifest: list[dict[str, Any]], args: argparse.Namespace) -> dict[str, Any]:
    query_idx = block_indices(manifest, include=["video_fisheye_cam0"])
    target_idx = block_indices(manifest, include=["video_stereo_left"])
    if len(query_idx) == 0 or len(target_idx) == 0:
        raise ValueError("Expected video_fisheye_cam0 and video_stereo_left feature blocks.")
    train, test = chronological_split(len(X), args.train_fraction)
    Xq = X[:, query_idx]
    Yt = X[:, target_idx]
    pred_min = ridge_predict(Xq[train], Yt[train], Xq[test], l2=args.ridge_l2, standardize_y=True)
    min_metrics, min_rows = retrieval_metrics(pred_min, Yt[test])
    for row in min_rows:
        idx = int(test[int(row["test_position"])])
        row["window_index"] = idx
        row["center_frame"] = rows[idx]["center_frame"]
    write_csv(OUT_DIR / "multi_view_consistency_minimal_ranks.csv", min_rows, list(min_rows[0].keys()))

    neural = {"available": False, "reason": "skipped by flag"}
    neural_metrics = None
    neural_rows: list[dict[str, Any]] = []
    if not args.skip_neural:
        pred_neural, neural = train_neural(Xq[train], Yt[train], Xq[test], task_type="projection", args=args)
        if neural.get("available") and pred_neural.size:
            neural_metrics, neural_rows = retrieval_metrics(pred_neural, Yt[test])
            for row in neural_rows:
                idx = int(test[int(row["test_position"])])
                row["window_index"] = idx
                row["center_frame"] = rows[idx]["center_frame"]
            write_csv(OUT_DIR / "multi_view_consistency_neural_ranks.csv", neural_rows, list(neural_rows[0].keys()))

    return {
        "train_windows": int(len(train)),
        "test_windows": int(len(test)),
        "query_block": "video_fisheye_cam0",
        "target_block": "video_stereo_left",
        "query_dim": int(len(query_idx)),
        "target_dim": int(len(target_idx)),
        "minimal": min_metrics,
        "neural_mlp": neural_metrics,
        "neural_training": neural,
    }


def task_action_phase_progress(X: np.ndarray, rows: list[dict[str, str]], manifest: list[dict[str, Any]], args: argparse.Namespace) -> dict[str, Any]:
    input_idx = block_indices(manifest, exclude=["caption_objects_interaction_text"])
    target = action_progress_targets(rows)
    train, test = chronological_split(len(X), args.train_fraction)
    pred_min = ridge_predict(X[train][:, input_idx], target[train], X[test][:, input_idx], l2=args.ridge_l2, standardize_y=True)[:, 0]
    pred_min = np.clip(pred_min, 0.0, 1.0)
    min_metrics = regression_metrics(target[test], pred_min)
    min_rows = []
    for local_pos, pred in zip(test, pred_min):
        idx = int(local_pos)
        min_rows.append(
            {
                "window_index": idx,
                "center_frame": rows[idx]["center_frame"],
                "action_label": rows[idx]["action_label"],
                "true_progress": float(target[idx]),
                "pred_progress": float(pred),
                "absolute_error": float(abs(pred - target[idx])),
            }
        )
    write_csv(OUT_DIR / "action_phase_progress_minimal_predictions.csv", min_rows, list(min_rows[0].keys()))

    neural = {"available": False, "reason": "skipped by flag"}
    neural_metrics = None
    neural_rows: list[dict[str, Any]] = []
    if not args.skip_neural:
        pred_neural, neural = train_neural(X[train][:, input_idx], target[train], X[test][:, input_idx], task_type="regression", args=args)
        if neural.get("available") and pred_neural.size:
            values = np.clip(pred_neural[:, 0], 0.0, 1.0)
            neural_metrics = regression_metrics(target[test], values)
            for local_pos, pred in zip(test, values):
                idx = int(local_pos)
                neural_rows.append(
                    {
                        "window_index": idx,
                        "center_frame": rows[idx]["center_frame"],
                        "action_label": rows[idx]["action_label"],
                        "true_progress": float(target[idx]),
                        "pred_progress": float(pred),
                        "absolute_error": float(abs(pred - target[idx])),
                    }
                )
            write_csv(OUT_DIR / "action_phase_progress_neural_predictions.csv", neural_rows, list(neural_rows[0].keys()))

    return {
        "train_windows": int(len(train)),
        "test_windows": int(len(test)),
        "input_dim": int(len(input_idx)),
        "target_source": "normalized position inside contiguous action-label runs",
        "minimal": min_metrics,
        "neural_mlp": neural_metrics,
        "neural_training": neural,
    }


def task_ego_motion_forecast(X: np.ndarray, rows: list[dict[str, str]], manifest: list[dict[str, Any]], args: argparse.Namespace) -> dict[str, Any]:
    input_idx = block_indices(manifest, exclude=["camera_translation", "caption_objects_interaction_text"])
    target_idx = block_indices(manifest, include=["camera_translation"])
    horizon = int(args.future_windows)
    valid = np.arange(0, len(X) - horizon, dtype=np.int64)
    target = X[valid + horizon][:, target_idx] - X[valid][:, target_idx]
    Xv = X[valid][:, input_idx]
    train, test = chronological_split(len(valid), args.train_fraction)
    pred_min = ridge_predict(Xv[train], target[train], Xv[test], l2=args.ridge_l2, standardize_y=True)
    min_metrics = regression_metrics(target[test], pred_min)
    min_rows = []
    for local_pos, pred in zip(test, pred_min):
        idx = int(valid[int(local_pos)])
        true_delta = target[int(local_pos)]
        min_rows.append(
            {
                "window_index": idx,
                "center_frame": rows[idx]["center_frame"],
                "future_window_index": int(idx + horizon),
                "delta_l2_true": float(np.linalg.norm(true_delta)),
                "delta_l2_pred": float(np.linalg.norm(pred)),
                "delta_l2_error": float(np.linalg.norm(pred - true_delta)),
            }
        )
    write_csv(OUT_DIR / "ego_motion_forecast_minimal_predictions.csv", min_rows, list(min_rows[0].keys()))

    neural = {"available": False, "reason": "skipped by flag"}
    neural_metrics = None
    neural_rows: list[dict[str, Any]] = []
    if not args.skip_neural:
        pred_neural, neural = train_neural(Xv[train], target[train], Xv[test], task_type="regression", args=args)
        if neural.get("available") and pred_neural.size:
            neural_metrics = regression_metrics(target[test], pred_neural)
            for local_pos, pred in zip(test, pred_neural):
                idx = int(valid[int(local_pos)])
                true_delta = target[int(local_pos)]
                neural_rows.append(
                    {
                        "window_index": idx,
                        "center_frame": rows[idx]["center_frame"],
                        "future_window_index": int(idx + horizon),
                        "delta_l2_true": float(np.linalg.norm(true_delta)),
                        "delta_l2_pred": float(np.linalg.norm(pred)),
                        "delta_l2_error": float(np.linalg.norm(pred - true_delta)),
                    }
                )
            write_csv(OUT_DIR / "ego_motion_forecast_neural_predictions.csv", neural_rows, list(neural_rows[0].keys()))

    return {
        "train_windows": int(len(train)),
        "test_windows": int(len(test)),
        "forecast_horizon_windows": horizon,
        "forecast_horizon_frames": int(horizon * 5),
        "input_dim": int(len(input_idx)),
        "target_dim": int(len(target_idx)),
        "target_source": "future minus current camera_translation feature block",
        "minimal": min_metrics,
        "neural_mlp": neural_metrics,
        "neural_training": neural,
    }


def fmt_metric(value: float | None, metric_key: str) -> str:
    if value is None:
        return "n/a"
    if metric_key in {"mae", "mse"}:
        return f"{value:.4f}"
    return f"{value:.4f}"


def task_main_metric(task: str, result: dict[str, Any], baseline: str) -> float | None:
    metrics = result.get(baseline)
    if not metrics:
        return None
    key = TASK_SPECS[task]["metric_key"]
    value = metrics.get(key)
    return float(value) if value is not None else None


def write_markdown(payload: dict[str, Any]) -> None:
    lines = [
        "# Four-Direction Extension Task Baselines",
        "",
        "Generated by `scripts/research_direction_extension_tasks.py` from the committed single-episode feature tensor.",
        "These are data-backed extension probes that show how each research direction can be started from Xperience-10M modalities.",
        "They do not claim cross-episode generalization or full direction completion.",
        "",
        "## Summary",
        "",
        "| Direction | Extension task | Minimal | Neural MLP | Meaning |",
        "| --- | --- | ---: | ---: | --- |",
    ]
    for task, spec in TASK_SPECS.items():
        result = payload["tasks"][task]
        key = spec["metric_key"]
        min_value = task_main_metric(task, result, "minimal")
        nn_value = task_main_metric(task, result, "neural_mlp")
        lines.append(
            f"| {spec['direction']}. {spec['direction_name']} | `{task}` | {fmt_metric(min_value, key)} {spec['metric_name']} | {fmt_metric(nn_value, key)} {spec['metric_name']} | {spec['current_limit']} |"
        )

    lines.extend(["", "## Task Details", ""])
    for task, spec in TASK_SPECS.items():
        result = payload["tasks"][task]
        key = spec["metric_key"]
        lines.extend(
            [
                f"### {spec['direction']}. {spec['name']}",
                "",
                f"- Case study: {spec['case_study']}",
                f"- Input: {spec['input']}",
                f"- Middle process modules: {spec['middle_process']}",
                f"- Output: {spec['output']}",
                f"- Minimal baseline: {spec['minimal_baseline']}",
                f"- Neural baseline: {spec['neural_baseline']}",
                f"- Minimal result: {fmt_metric(task_main_metric(task, result, 'minimal'), key)} {spec['metric_name']}",
                f"- Neural result: {fmt_metric(task_main_metric(task, result, 'neural_mlp'), key)} {spec['metric_name']}",
                f"- Limitation: {spec['current_limit']}",
                "",
            ]
        )

    (OUT_DIR / "research_direction_extension_summary.md").write_text("\n".join(lines).rstrip() + "\n", encoding="utf-8")


def svg_text(x: int, y: int, text: str, size: int = 16, weight: int = 500, color: str = "#f4f8ef") -> str:
    return (
        f'<text x="{x}" y="{y}" font-size="{size}" font-weight="{weight}" '
        f'fill="{color}">{html.escape(text)}</text>'
    )


def write_svg(payload: dict[str, Any]) -> None:
    CHARTS.mkdir(parents=True, exist_ok=True)
    width = 1420
    height = 920
    colors = {"A": "#a7f078", "B": "#7ae5c3", "C": "#d8f4a5", "D": "#9bdfff"}
    svg: list[str] = [
        f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}">',
        '<rect width="1420" height="920" fill="#020502"/>',
        '<rect x="28" y="28" width="1364" height="864" rx="18" fill="#050905" stroke="#a7f078" stroke-opacity="0.24"/>',
        svg_text(66, 88, "Ropedia Xperience-10M: four direction extension probes", 32, 760),
        svg_text(66, 122, "Data-backed from the same 1,161-window public sample feature tensor; extension probes, not full direction claims.", 17, 500, "#a5afa2"),
    ]
    x0 = 66
    y0 = 166
    card_w = 620
    card_h = 160
    gap_x = 44
    gap_y = 34
    for i, (task, spec) in enumerate(TASK_SPECS.items()):
        result = payload["tasks"][task]
        row = i // 2
        col = i % 2
        x = x0 + col * (card_w + gap_x)
        y = y0 + row * (card_h + gap_y)
        color = colors[spec["direction"]]
        min_v = task_main_metric(task, result, "minimal")
        nn_v = task_main_metric(task, result, "neural_mlp")
        metric = spec["metric_name"]
        svg.extend(
            [
                f'<rect x="{x}" y="{y}" width="{card_w}" height="{card_h}" rx="10" fill="#071207" stroke="#a7f078" stroke-opacity="0.22"/>',
                f'<rect x="{x}" y="{y}" width="10" height="{card_h}" rx="5" fill="{color}"/>',
                f'<circle cx="{x + 42}" cy="{y + 40}" r="24" fill="{color}" opacity="0.14"/>',
                svg_text(x + 32, y + 48, spec["direction"], 21, 760, color),
                svg_text(x + 76, y + 35, spec["name"], 20, 760),
                svg_text(x + 76, y + 62, spec["direction_name"], 13, 650, "#a5afa2"),
                svg_text(x + 76, y + 94, f"Minimal: {fmt_metric(min_v, spec['metric_key'])} {metric}", 16, 700, "#f4f8ef"),
                svg_text(x + 300, y + 94, f"Neural MLP: {fmt_metric(nn_v, spec['metric_key'])} {metric}", 16, 700, "#f4f8ef"),
                svg_text(x + 76, y + 125, spec["output"], 13, 500, "#dce8d7"),
            ]
        )
        min_score = choose_score(task, result["minimal"])
        nn_score = choose_score(task, result["neural_mlp"]) if result.get("neural_mlp") else 0.0
        bar_x = x + 76
        bar_y = y + 138
        bar_w = 440
        svg.append(f'<rect x="{bar_x}" y="{bar_y}" width="{bar_w}" height="8" rx="4" fill="#a7f078" opacity="0.14"/>')
        svg.append(f'<rect x="{bar_x}" y="{bar_y}" width="{max(4, min(bar_w, bar_w * min_score)):.1f}" height="8" rx="4" fill="{color}" opacity="0.72"/>')
        svg.append(f'<rect x="{bar_x}" y="{bar_y + 12}" width="{bar_w}" height="8" rx="4" fill="#a7f078" opacity="0.14"/>')
        svg.append(f'<rect x="{bar_x}" y="{bar_y + 12}" width="{max(4, min(bar_w, bar_w * nn_score)):.1f}" height="8" rx="4" fill="#ffffff" opacity="0.78"/>')

    legend_y = 570
    svg.extend(
        [
            svg_text(66, legend_y, "How to read this", 24, 760),
            svg_text(66, legend_y + 34, "Each card adds one concrete task to a research direction using existing sample modalities.", 16, 500, "#dce8d7"),
            svg_text(66, legend_y + 62, "Colored bar: minimal baseline normalized score. White bar: neural MLP normalized score. Lower-is-better MAE is shown as 1 - MAE for bar length only.", 16, 500, "#dce8d7"),
            '<line x1="66" y1="675" x2="1354" y2="675" stroke="#a7f078" stroke-opacity="0.18"/>',
            svg_text(66, 724, "Implementation boundary", 22, 760),
            svg_text(66, 758, "A: motion-energy proxy, not a full human body model. B: view-feature retrieval, not neural rendering.", 16, 500, "#dce8d7"),
            svg_text(66, 786, "C: phase-progress regression, not open-world intent. D: ego-motion forecast, not a persistent map.", 16, 500, "#dce8d7"),
            svg_text(66, 835, "All metrics are computed from held-out chronological windows of the same public sample episode.", 16, 700, "#f4f8ef"),
        ]
    )
    svg.append("</svg>")
    (CHARTS / "research_direction_extension_tasks.svg").write_text("\n".join(svg), encoding="utf-8")


def build_payload(args: argparse.Namespace) -> dict[str, Any]:
    X, starts, ends, rows, manifest = load_inputs(args.results_dir)
    global OUT_DIR
    OUT_DIR = args.output_dir
    OUT_DIR.mkdir(parents=True, exist_ok=True)

    tasks = OrderedDict()
    tasks["body_motion_intensity"] = task_body_motion_intensity(X, rows, manifest, args)
    tasks["multi_view_consistency_retrieval"] = task_multi_view_retrieval(X, rows, manifest, args)
    tasks["action_phase_progress"] = task_action_phase_progress(X, rows, manifest, args)
    tasks["ego_motion_forecast"] = task_ego_motion_forecast(X, rows, manifest, args)

    payload = {
        "source": {
            "shared_windows": str((args.results_dir / "shared_windows.npz").relative_to(ROOT)),
            "windows_csv": str((args.results_dir / "windows.csv").relative_to(ROOT)),
            "feature_manifest": str((args.results_dir / "feature_manifest.json").relative_to(ROOT)),
        },
        "dataset_scope": {
            "sample_episode_count": 1,
            "num_windows": int(len(X)),
            "feature_dim": int(X.shape[1]),
            "first_start_frame": int(starts[0]),
            "last_end_frame": int(ends[-1]),
            "warning": "Single public sample episode; these extension probes validate task design and pipeline mechanics, not cross-episode generalization.",
        },
        "baselines": {
            "minimal": "Ridge classifiers/regressors/projections plus cosine retrieval on the committed feature tensor.",
            "neural_mlp": "Small one-hidden-layer PyTorch MLP heads using the same inputs, targets, chronological split, and evaluator.",
        },
        "run_config": {
            "train_fraction": float(args.train_fraction),
            "ridge_l2": float(args.ridge_l2),
            "seed": int(args.seed),
            "future_windows": int(args.future_windows),
            "neural_epochs": int(args.neural_epochs),
            "neural_hidden_dim": int(args.neural_hidden_dim),
            "neural_batch_size": int(args.neural_batch_size),
            "skip_neural": bool(args.skip_neural),
        },
        "task_specs": TASK_SPECS,
        "tasks": tasks,
    }
    return payload


def main() -> int:
    args = parse_args()
    payload = build_payload(args)
    write_json(args.output_dir / "research_direction_extension_results.json", payload)
    write_json(DOCS_DATA / "research_direction_extensions.json", payload)
    write_markdown(payload)
    write_svg(payload)
    print(f"Wrote {args.output_dir / 'research_direction_extension_results.json'}")
    print(f"Wrote {CHARTS / 'research_direction_extension_tasks.svg'}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())