exoskeleton/code/analysis/metrics.py

"""Independent H1--H4 trial metrics.

These functions consume stored arrays only.  They deliberately do not import
the simulation or controller implementations, so paper endpoints can be
reconstructed independently from raw evidence.
"""

from __future__ import annotations

from typing import Any, Mapping

import numpy as np


METRIC_SCHEMA_VERSION = "1.1.0"


class MetricError(ValueError):
    pass


def _vector(value: Any, name: str, *, dtype=float) -> np.ndarray:
    array = np.asarray(value, dtype=dtype)
    if array.ndim != 1 or array.size == 0:
        raise MetricError(f"{name} must be a non-empty one-dimensional array")
    return array


def _matrix(value: Any, name: str, columns: int | None = None) -> np.ndarray:
    array = np.asarray(value, dtype=float)
    if array.ndim != 2 or array.shape[0] == 0:
        raise MetricError(f"{name} must be a non-empty two-dimensional array")
    if columns is not None and array.shape[1] != columns:
        raise MetricError(f"{name} must have {columns} columns")
    return array


def _same_rows(named: Mapping[str, np.ndarray]) -> int:
    counts = {name: value.shape[0] for name, value in named.items()}
    if len(set(counts.values())) != 1:
        raise MetricError(f"sample counts differ: {counts}")
    return next(iter(counts.values()))


def _finite(array: np.ndarray, name: str) -> None:
    if not np.all(np.isfinite(array)):
        raise MetricError(f"{name} contains non-finite values")


def _wrapped_delta(array: np.ndarray) -> np.ndarray:
    return (array + np.pi) % (2.0 * np.pi) - np.pi


def compute_h1_composite(
    *,
    mapping_valid: Any,
    position_error_m: Any,
    orientation_error_rad: Any,
    q_slave: Any,
    swivel_angle_rad: Any,
    master_step_norm: Any,
    position_threshold_m: float,
    orientation_threshold_rad: float,
    joint_step_threshold_rad: float,
    swivel_step_threshold_rad: float,
    input_step_threshold_rad: float,
    accepted: Any | None = None,
    commanded_reset: Any | None = None,
    degeneracy_transition: Any | None = None,
) -> dict[str, Any]:
    """Compute the trajectory-level H1 ``F_r``, ``D_r``, and ``C_r``."""
    valid = _vector(mapping_valid, "mapping_valid", dtype=bool)
    e_position = _vector(position_error_m, "position_error_m")
    e_orientation = _vector(orientation_error_rad, "orientation_error_rad")
    slave = _matrix(q_slave, "q_slave")
    swivel = _vector(swivel_angle_rad, "swivel_angle_rad")
    input_step = _vector(master_step_norm, "master_step_norm")
    n = _same_rows(
        {
            "mapping_valid": valid,
            "position_error_m": e_position,
            "orientation_error_rad": e_orientation,
            "q_slave": slave,
            "swivel_angle_rad": swivel,
            "master_step_norm": input_step,
        }
    )
    accepted_array = (
        np.ones(n, dtype=bool)
        if accepted is None
        else _vector(accepted, "accepted", dtype=bool)
    )
    reset = (
        np.zeros(n, dtype=bool)
        if commanded_reset is None
        else _vector(commanded_reset, "commanded_reset", dtype=bool)
    )
    degeneracy = (
        np.zeros(n, dtype=bool)
        if degeneracy_transition is None
        else _vector(
            degeneracy_transition,
            "degeneracy_transition",
            dtype=bool,
        )
    )
    _same_rows(
        {
            "accepted": accepted_array,
            "commanded_reset": reset,
            "degeneracy_transition": degeneracy,
            "mapping_valid": valid,
        }
    )
    for array, name in (
        (e_position, "position_error_m"),
        (e_orientation, "orientation_error_rad"),
        (slave, "q_slave"),
        (swivel, "swivel_angle_rad"),
        (input_step, "master_step_norm"),
    ):
        _finite(array, name)
    thresholds = {
        "position_threshold_m": position_threshold_m,
        "orientation_threshold_rad": orientation_threshold_rad,
        "joint_step_threshold_rad": joint_step_threshold_rad,
        "swivel_step_threshold_rad": swivel_step_threshold_rad,
        "input_step_threshold_rad": input_step_threshold_rad,
    }
    if any(not np.isfinite(value) or value < 0.0 for value in thresholds.values()):
        raise MetricError("H1 thresholds must be finite and non-negative")

    failure_mask = accepted_array & (
        ~valid
        | (e_position > position_threshold_m)
        | (e_orientation > orientation_threshold_rad)
    )
    if n > 1:
        joint_step = np.max(
            np.abs(_wrapped_delta(slave[1:] - slave[:-1])),
            axis=1,
        )
        swivel_step = np.abs(_wrapped_delta(swivel[1:] - swivel[:-1]))
        eligible = (
            accepted_array[1:]
            & accepted_array[:-1]
            & valid[1:]
            & valid[:-1]
            & (input_step[1:] <= input_step_threshold_rad)
            & ~reset[1:]
            & ~degeneracy[1:]
        )
        discontinuity_mask = eligible & (
            (joint_step > joint_step_threshold_rad)
            | (swivel_step > swivel_step_threshold_rad)
        )
    else:
        joint_step = np.empty(0, dtype=float)
        swivel_step = np.empty(0, dtype=float)
        eligible = np.empty(0, dtype=bool)
        discontinuity_mask = np.empty(0, dtype=bool)

    F_r = int(np.any(failure_mask))
    D_r = int(np.any(discontinuity_mask))
    return {
        "h1_F_r": F_r,
        "h1_D_r": D_r,
        "h1_C_r": max(F_r, D_r),
        "h1_failure_sample_count": int(np.sum(failure_mask)),
        "h1_discontinuity_sample_count": int(np.sum(discontinuity_mask)),
        "h1_eligible_increment_count": int(np.sum(eligible)),
        "h1_mapping_valid_fraction": float(np.mean(valid[accepted_array]))
        if np.any(accepted_array)
        else 0.0,
        "h1_max_position_error_m": float(np.max(e_position[accepted_array]))
        if np.any(accepted_array)
        else 0.0,
        "h1_max_orientation_error_rad": float(np.max(e_orientation[accepted_array]))
        if np.any(accepted_array)
        else 0.0,
        "h1_max_joint_step_rad": float(np.max(joint_step))
        if joint_step.size
        else 0.0,
        "h1_max_swivel_step_rad": float(np.max(swivel_step))
        if swivel_step.size
        else 0.0,
    }


def compute_h1_audit_metrics(
    *,
    differential_valid: Any,
    validity_reason_code: Any,
    reach_clip_code: Any,
    joint_limit_active: Any,
    geometry_degenerate: Any,
    low_manipulability: Any,
    slave_min_singular_value: Any,
    slave_manipulability: Any,
    validity_reason_labels: Any | None = None,
) -> dict[str, Any]:
    """Summarize H1 stratum isolation and invalid-sample explanations."""
    differential = _vector(
        differential_valid, "differential_valid", dtype=bool
    )
    reason_raw = _vector(validity_reason_code, "validity_reason_code")
    reach_raw = _vector(reach_clip_code, "reach_clip_code")
    joint_limit = _vector(
        joint_limit_active, "joint_limit_active", dtype=bool
    )
    geometry = _vector(
        geometry_degenerate, "geometry_degenerate", dtype=bool
    )
    low_manip = _vector(
        low_manipulability, "low_manipulability", dtype=bool
    )
    minimum_singular = _vector(
        slave_min_singular_value, "slave_min_singular_value"
    )
    manipulability = _vector(slave_manipulability, "slave_manipulability")
    _same_rows(
        {
            "differential_valid": differential,
            "validity_reason_code": reason_raw,
            "reach_clip_code": reach_raw,
            "joint_limit_active": joint_limit,
            "geometry_degenerate": geometry,
            "low_manipulability": low_manip,
            "slave_min_singular_value": minimum_singular,
            "slave_manipulability": manipulability,
        }
    )
    _finite(reason_raw, "validity_reason_code")
    _finite(reach_raw, "reach_clip_code")
    _finite(minimum_singular, "slave_min_singular_value")
    _finite(manipulability, "slave_manipulability")
    if np.any(minimum_singular < 0.0) or np.any(manipulability < 0.0):
        raise MetricError(
            "H1 singular-value and manipulability diagnostics must be non-negative"
        )
    if not np.all(reason_raw == np.floor(reason_raw)) or np.any(reason_raw < 0):
        raise MetricError("validity_reason_code must contain non-negative integers")
    if not np.all(reach_raw == np.floor(reach_raw)) or not set(
        np.asarray(reach_raw, dtype=int).tolist()
    ).issubset({-1, 0, 1}):
        raise MetricError("reach_clip_code must contain only -1, 0, or 1")
    reason = np.asarray(reason_raw, dtype=int)
    reach = np.asarray(reach_raw, dtype=int)

    if validity_reason_labels is None:
        labels: list[str] = []
    elif (
        not isinstance(validity_reason_labels, list)
        or any(not isinstance(label, str) or not label for label in validity_reason_labels)
    ):
        raise MetricError("H1 validity_reason_labels must be a list of strings")
    else:
        labels = list(validity_reason_labels)
    maximum_code = int(np.max(reason))
    if labels and maximum_code >= len(labels):
        raise MetricError(
            "H1 validity_reason_labels does not cover every recorded code"
        )

    unique_codes, counts = np.unique(reason, return_counts=True)
    histogram = {
        (labels[int(code)] if labels else f"code_{int(code)}"): int(count)
        for code, count in zip(unique_codes, counts)
    }
    invalid = ~differential
    invalid_codes = reason[invalid]
    if invalid_codes.size:
        invalid_unique, invalid_counts = np.unique(
            invalid_codes, return_counts=True
        )
        highest = int(np.max(invalid_counts))
        # Ties use the lowest stable enum code.
        primary_code = int(
            np.min(invalid_unique[invalid_counts == highest])
        )
        primary_reason = (
            labels[primary_code] if labels else f"code_{primary_code}"
        )
    else:
        primary_reason = labels[0] if labels else "code_0"

    return {
        "h1_reach_clip_fraction": float(np.mean(reach != 0)),
        "h1_reach_clip_lower_fraction": float(np.mean(reach == -1)),
        "h1_reach_clip_upper_fraction": float(np.mean(reach == 1)),
        "h1_low_manipulability_fraction": float(np.mean(low_manip)),
        "h1_joint_limit_active_fraction": float(np.mean(joint_limit)),
        "h1_geometry_degenerate_fraction": float(np.mean(geometry)),
        "h1_min_slave_min_singular_value": float(np.min(minimum_singular)),
        "h1_min_slave_manipulability": float(np.min(manipulability)),
        "h1_validity_reason_histogram": histogram,
        "h1_primary_invalid_reason": primary_reason,
        "h1_invalid_reason_sample_count": int(invalid_codes.size),
        "h1_unexplained_invalid_fraction": float(
            np.mean(invalid & (reason == 0))
        ),
    }


def compute_h2_wrench_metrics(
    wrench_estimated: Any,
    wrench_reference: Any,
    *,
    sample_mask: Any | None = None,
    numerical_rank_deficient: Any | None = None,
    operationally_ill_conditioned: Any | None = None,
    numerical_rank_threshold: Any | None = None,
    operational_min_scaled_singular_threshold: Any | None = None,
    scaled_singular_values: Any | None = None,
    condition_number: Any | None = None,
) -> dict[str, Any]:
    estimate = _matrix(wrench_estimated, "wrench_estimated", columns=6)
    reference = _matrix(wrench_reference, "wrench_reference", columns=6)
    n = _same_rows({"wrench_estimated": estimate, "wrench_reference": reference})
    _finite(estimate, "wrench_estimated")
    _finite(reference, "wrench_reference")
    mask = (
        np.ones(n, dtype=bool)
        if sample_mask is None
        else _vector(sample_mask, "sample_mask", dtype=bool)
    )
    if mask.shape[0] != n or not np.any(mask):
        raise MetricError("H2 sample_mask must select at least one aligned sample")
    error = estimate[mask] - reference[mask]
    force_norm = np.linalg.norm(error[:, :3], axis=1)
    moment_norm = np.linalg.norm(error[:, 3:], axis=1)
    result = {
        "h2_sample_count": int(error.shape[0]),
        "h2_force_rmse_N": float(np.sqrt(np.mean(force_norm**2))),
        "h2_moment_rmse_Nm": float(np.sqrt(np.mean(moment_norm**2))),
        "h2_force_mae_N": float(np.mean(force_norm)),
        "h2_moment_mae_Nm": float(np.mean(moment_norm)),
        "h2_force_bias_xyz_N": np.mean(error[:, :3], axis=0).tolist(),
        "h2_moment_bias_xyz_Nm": np.mean(error[:, 3:], axis=0).tolist(),
    }
    if numerical_rank_deficient is not None:
        numerical_flag = _vector(
            numerical_rank_deficient,
            "numerical_rank_deficient",
            dtype=bool,
        )
        if numerical_flag.shape[0] != n:
            raise MetricError("H2 numerical-rank flag length mismatch")
        result["h2_numerical_rank_deficient_fraction"] = float(
            np.mean(numerical_flag[mask])
        )
    if operationally_ill_conditioned is not None:
        operational_flag = _vector(
            operationally_ill_conditioned,
            "operationally_ill_conditioned",
            dtype=bool,
        )
        if operational_flag.shape[0] != n:
            raise MetricError("H2 operational-condition flag length mismatch")
        result["h2_operationally_ill_conditioned_fraction"] = float(
            np.mean(operational_flag[mask])
        )
    if numerical_rank_threshold is not None:
        numerical_threshold = _vector(
            numerical_rank_threshold,
            "numerical_rank_threshold",
        )
        if numerical_threshold.shape[0] != n:
            raise MetricError("H2 numerical-rank threshold length mismatch")
        _finite(numerical_threshold, "numerical_rank_threshold")
        result["h2_numerical_rank_threshold_min"] = float(
            np.min(numerical_threshold[mask])
        )
        result["h2_numerical_rank_threshold_max"] = float(
            np.max(numerical_threshold[mask])
        )
    if operational_min_scaled_singular_threshold is not None:
        operational_threshold = _vector(
            operational_min_scaled_singular_threshold,
            "operational_min_scaled_singular_threshold",
        )
        if operational_threshold.shape[0] != n:
            raise MetricError("H2 operational threshold length mismatch")
        _finite(
            operational_threshold,
            "operational_min_scaled_singular_threshold",
        )
        selected_threshold = operational_threshold[mask]
        if not np.allclose(
            selected_threshold,
            selected_threshold[0],
            rtol=0.0,
            atol=1e-15,
        ):
            raise MetricError(
                "H2 operational threshold must be constant within a trial"
            )
        result["h2_operational_min_scaled_singular_threshold"] = float(
            selected_threshold[0]
        )
    if scaled_singular_values is not None:
        singular_values = _matrix(
            scaled_singular_values,
            "scaled_singular_values",
        )
        if singular_values.shape[0] != n:
            raise MetricError("H2 singular-value sample count mismatch")
        _finite(singular_values, "scaled_singular_values")
        result["h2_min_scaled_singular_value"] = float(
            np.min(singular_values[mask, -1])
        )
    if condition_number is not None:
        condition = _vector(condition_number, "condition_number")
        if condition.shape[0] != n:
            raise MetricError("H2 condition-number sample count mismatch")
        if np.any(np.isnan(condition)) or np.any(condition < 0.0):
            raise MetricError(
                "H2 condition number must be non-negative and not NaN"
            )
        selected_condition = condition[mask]
        finite_condition = selected_condition[np.isfinite(selected_condition)]
        result["h2_infinite_condition_number_fraction"] = float(
            np.mean(~np.isfinite(selected_condition))
        )
        result["h2_max_finite_condition_number"] = (
            float(np.max(finite_condition))
            if finite_condition.size
            else None
        )
        # A rank-deficient sample has infinite condition number.  JSON cannot
        # represent infinity, so the explicit fraction above carries that
        # condition while this field is null in that case.
        result["h2_max_condition_number"] = (
            float(np.max(selected_condition))
            if np.all(np.isfinite(selected_condition))
            else None
        )
    return result


def _dt_array(dt: Any, count: int) -> np.ndarray:
    array = np.asarray(dt, dtype=float)
    if array.ndim == 0:
        array = np.full(count, float(array), dtype=float)
    if array.shape != (count,):
        raise MetricError(f"dt must be scalar or have shape ({count},)")
    if not np.all(np.isfinite(array)) or np.any(array <= 0.0):
        raise MetricError("dt must contain finite positive intervals")
    return array


def compute_h3_power_mismatch(
    *,
    tau_master_raw: Any,
    qd_master: Any,
    tau_slave_source: Any,
    qd_slave_source: Any,
    dt: Any,
    force_scale: float = 1.0,
    epsilon_energy_J: float = 1e-12,
    minimum_power_activity_J: float = 0.0,
    return_valid: Any | None = None,
) -> dict[str, Any]:
    tau_m = _matrix(tau_master_raw, "tau_master_raw")
    qd_m = _matrix(qd_master, "qd_master")
    tau_s = _matrix(tau_slave_source, "tau_slave_source")
    qd_s = _matrix(qd_slave_source, "qd_slave_source")
    n = _same_rows(
        {
            "tau_master_raw": tau_m,
            "qd_master": qd_m,
            "tau_slave_source": tau_s,
            "qd_slave_source": qd_s,
        }
    )
    if tau_m.shape != qd_m.shape or tau_s.shape != qd_s.shape:
        raise MetricError("torque and velocity shapes must match at each port")
    for array, name in (
        (tau_m, "tau_master_raw"),
        (qd_m, "qd_master"),
        (tau_s, "tau_slave_source"),
        (qd_s, "qd_slave_source"),
    ):
        _finite(array, name)
    intervals = _dt_array(dt, n)
    chi = (
        np.ones(n, dtype=bool)
        if return_valid is None
        else _vector(return_valid, "return_valid", dtype=bool)
    )
    if chi.shape[0] != n:
        raise MetricError("return_valid length mismatch")
    force_scale = float(force_scale)
    epsilon_energy_J = float(epsilon_energy_J)
    minimum_power_activity_J = float(minimum_power_activity_J)
    if not np.isfinite(force_scale) or force_scale < 0.0:
        raise MetricError("force_scale must be finite and non-negative")
    if not np.isfinite(epsilon_energy_J) or epsilon_energy_J <= 0.0:
        raise MetricError("epsilon_energy_J must be finite and positive")
    if (
        not np.isfinite(minimum_power_activity_J)
        or minimum_power_activity_J < 0.0
    ):
        raise MetricError(
            "minimum_power_activity_J must be finite and non-negative"
        )

    power_master = np.einsum("ij,ij->i", tau_m, qd_m)
    power_slave = chi.astype(float) * np.einsum("ij,ij->i", tau_s, qd_s)
    scaled_slave = force_scale * power_slave
    numerator = float(np.sum(np.abs(power_master - scaled_slave) * intervals))
    power_activity = float(
        0.5
        * np.sum((np.abs(power_master) + np.abs(scaled_slave)) * intervals)
    )
    denominator = power_activity + epsilon_energy_J
    normalized = numerator / denominator
    normalized_valid = bool(power_activity >= minimum_power_activity_J)
    return {
        # Kept for backward-compatible diagnostics.  Confirmatory analysis must
        # use the validity flag/gated value when a nonzero activity gate is set.
        "h3_epsilon_P_act": normalized,
        "h3_epsilon_P_act_gated": normalized if normalized_valid else None,
        "h3_normalized_metric_valid": normalized_valid,
        "h3_minimum_power_activity_J": minimum_power_activity_J,
        "h3_power_activity_J": power_activity,
        "h3_absolute_power_mismatch_J": numerator,
        "h3_power_mismatch_numerator_J": numerator,
        "h3_power_normalizer_J": denominator,
        "h3_return_valid_fraction": float(np.mean(chi)),
        "h3_master_raw_work_J": float(np.sum(power_master * intervals)),
        "h3_scaled_slave_work_J": float(np.sum(scaled_slave * intervals)),
    }


def audit_h4_energy(
    *,
    energy_before_J: Any,
    energy_after_J: Any,
    tau_candidate: Any,
    tau_applied: Any,
    tau_accepted: Any | None = None,
    qd_master: Any,
    dt: Any,
    energy_min_J: float,
    energy_max_J: float,
    epsilon_torque_impulse_Nms: float = 1e-12,
    audit_tolerance_J: float = 1e-10,
    software_preclip_J: Any | None = None,
) -> dict[str, Any]:
    energy_before = _vector(energy_before_J, "energy_before_J")
    energy_after = _vector(energy_after_J, "energy_after_J")
    candidate = _matrix(tau_candidate, "tau_candidate")
    applied = _matrix(tau_applied, "tau_applied")
    accepted = (
        applied
        if tau_accepted is None
        else _matrix(tau_accepted, "tau_accepted")
    )
    velocity = _matrix(qd_master, "qd_master")
    n = _same_rows(
        {
            "energy_before_J": energy_before,
            "energy_after_J": energy_after,
            "tau_candidate": candidate,
            "tau_applied": applied,
            "tau_accepted": accepted,
            "qd_master": velocity,
        }
    )
    if not (
        candidate.shape
        == applied.shape
        == accepted.shape
        == velocity.shape
    ):
        raise MetricError(
            "candidate, projected, accepted torque, and velocity must align"
        )
    for array, name in (
        (energy_before, "energy_before_J"),
        (energy_after, "energy_after_J"),
        (candidate, "tau_candidate"),
        (applied, "tau_applied"),
        (accepted, "tau_accepted"),
        (velocity, "qd_master"),
    ):
        _finite(array, name)
    intervals = _dt_array(dt, n)
    energy_min_J = float(energy_min_J)
    energy_max_J = float(energy_max_J)
    tolerance = float(audit_tolerance_J)
    epsilon_tau = float(epsilon_torque_impulse_Nms)
    if not (
        np.isfinite(energy_min_J)
        and np.isfinite(energy_max_J)
        and 0.0 <= energy_min_J <= energy_max_J
    ):
        raise MetricError("invalid energy bounds")
    if not np.isfinite(tolerance) or tolerance < 0.0:
        raise MetricError("audit_tolerance_J must be finite and non-negative")
    if not np.isfinite(epsilon_tau) or epsilon_tau <= 0.0:
        raise MetricError(
            "epsilon_torque_impulse_Nms must be finite and positive"
        )

    # The deterministic gate is reconstructed at the actuator-accepted port.
    # If no drive readback exists, callers may omit tau_accepted and explicitly
    # declare that projected == accepted for that backend.
    applied_power = np.einsum("ij,ij->i", accepted, velocity)
    candidate_power = np.einsum("ij,ij->i", candidate, velocity)
    reconstructed_preclip = energy_before - applied_power * intervals
    reconstructed_after = np.clip(
        reconstructed_preclip,
        energy_min_J,
        energy_max_J,
    )
    deficit = np.maximum(0.0, energy_min_J - reconstructed_preclip)
    accounting_error = np.abs(energy_after - reconstructed_after)

    shadow_energy = float(energy_before[0])
    shadow_min = shadow_energy
    for power, interval in zip(candidate_power, intervals):
        shadow_energy = min(energy_max_J, shadow_energy - float(power * interval))
        shadow_min = min(shadow_min, shadow_energy)
    shadow_deficit = max(0.0, energy_min_J - shadow_min)
    projected_deficit = float(np.max(deficit))

    numerator = float(
        np.sum(np.linalg.norm(accepted - candidate, axis=1) * intervals)
    )
    denominator = float(
        np.sum(np.linalg.norm(candidate, axis=1) * intervals) + epsilon_tau
    )
    preclip_discrepancy = 0.0
    if software_preclip_J is not None:
        software = _vector(software_preclip_J, "software_preclip_J")
        if software.shape[0] != n:
            raise MetricError("software_preclip_J length mismatch")
        _finite(software, "software_preclip_J")
        preclip_discrepancy = float(
            np.max(np.abs(software - reconstructed_preclip))
        )
    audit_pass = (
        projected_deficit <= tolerance
        and float(np.max(accounting_error)) <= tolerance
        and preclip_discrepancy <= tolerance
    )
    return {
        "h4_energy_audit_pass": bool(audit_pass),
        "h4_preclip_floor_deficit_max_J": projected_deficit,
        "h4_energy_accounting_max_error_J": float(np.max(accounting_error)),
        "h4_software_preclip_max_error_J": preclip_discrepancy,
        "h4_shadow_floor_deficit_J": float(shadow_deficit),
        "h4_projected_floor_deficit_J": projected_deficit,
        "h4_delta_B_J": float(shadow_deficit - projected_deficit),
        "h4_D_proj": numerator / denominator,
        "h4_energy_min_J": energy_min_J,
        "h4_energy_max_J": energy_max_J,
        "h4_projection_distortion_numerator_Nms": numerator,
        "h4_projection_distortion_normalizer_Nms": denominator,
        "h4_shadow_energy_min_J": float(shadow_min),
        "h4_downstream_modification_max_Nm": float(
            np.max(np.linalg.norm(accepted - applied, axis=1))
        ),
    }


def compute_bilateral_diagnostics(
    *,
    master_tracking_error_rad: Any,
    slave_tracking_error_rad: Any,
    feedback_torque_Nm: Any,
    contact_force_N: Any,
    projection_factor: Any,
    projection_tolerance: float = 1e-12,
    contact_force_threshold_N: float = 1e-6,
) -> dict[str, Any]:
    """Compute secondary task/transparency diagnostics from stored samples."""
    master_error = _vector(
        master_tracking_error_rad, "master_tracking_error_rad"
    )
    slave_error = _vector(
        slave_tracking_error_rad, "slave_tracking_error_rad"
    )
    feedback = _matrix(feedback_torque_Nm, "feedback_torque_Nm")
    contact = _vector(contact_force_N, "contact_force_N")
    rho = _vector(projection_factor, "projection_factor")
    _same_rows(
        {
            "master_tracking_error_rad": master_error,
            "slave_tracking_error_rad": slave_error,
            "feedback_torque_Nm": feedback,
            "contact_force_N": contact,
            "projection_factor": rho,
        }
    )
    for array, name in (
        (master_error, "master_tracking_error_rad"),
        (slave_error, "slave_tracking_error_rad"),
        (feedback, "feedback_torque_Nm"),
        (contact, "contact_force_N"),
        (rho, "projection_factor"),
    ):
        _finite(array, name)
    projection_tolerance = float(projection_tolerance)
    contact_force_threshold_N = float(contact_force_threshold_N)
    if not np.isfinite(projection_tolerance) or projection_tolerance < 0.0:
        raise MetricError("projection_tolerance must be finite and non-negative")
    if (
        not np.isfinite(contact_force_threshold_N)
        or contact_force_threshold_N < 0.0
    ):
        raise MetricError(
            "contact_force_threshold_N must be finite and non-negative"
        )
    feedback_norm = np.linalg.norm(feedback, axis=1)
    return {
        "bilateral_master_tracking_rmse_rad": float(
            np.sqrt(np.mean(np.square(master_error)))
        ),
        "bilateral_slave_tracking_rmse_rad": float(
            np.sqrt(np.mean(np.square(slave_error)))
        ),
        "bilateral_feedback_torque_rms_Nm": float(
            np.sqrt(np.mean(np.square(feedback_norm)))
        ),
        "bilateral_feedback_torque_peak_Nm": float(np.max(feedback_norm)),
        "bilateral_contact_force_rms_N": float(
            np.sqrt(np.mean(np.square(contact)))
        ),
        "bilateral_contact_force_peak_N": float(np.max(contact)),
        "bilateral_contact_fraction": float(
            np.mean(contact > contact_force_threshold_N)
        ),
        "bilateral_projection_intervention_fraction": float(
            np.mean(rho < (1.0 - projection_tolerance))
        ),
        "bilateral_projection_factor_min": float(np.min(rho)),
    }


def _field(
    samples: Mapping[str, Any],
    fields: Mapping[str, str],
    logical_name: str,
    default_name: str,
    *,
    optional: bool = False,
) -> Any:
    stored_name = fields.get(logical_name, default_name)
    if stored_name not in samples:
        if optional:
            return None
        raise MetricError(
            f"missing sample field {stored_name!r} for {logical_name!r}"
        )
    return samples[stored_name]


def _constant_trial_value(value: Any, name: str) -> float:
    """Resolve a scalar or constant non-empty sample vector."""
    array = np.asarray(value, dtype=float)
    if array.ndim == 0:
        result = float(array)
    elif array.ndim == 1 and array.size > 0:
        _finite(array, name)
        result = float(array[0])
        if not np.allclose(array, result, rtol=0.0, atol=1e-15):
            raise MetricError(f"{name} must be constant within one trial")
    else:
        raise MetricError(f"{name} must be a scalar or non-empty vector")
    if not np.isfinite(result):
        raise MetricError(f"{name} must be finite")
    return result


def _energy_bound_for_trial(
    samples: Mapping[str, Any],
    fields: Mapping[str, str],
    family_config: Mapping[str, Any],
    logical_name: str,
    default_field: str,
) -> float:
    """Prefer stored effective bounds and reject config/sample disagreement."""
    stored = _field(
        samples,
        fields,
        logical_name,
        default_field,
        optional=True,
    )
    configured = family_config.get(logical_name)
    if stored is None and configured is None:
        raise MetricError(
            f"H4 requires {logical_name} in samples or metric configuration"
        )
    if stored is None:
        return _constant_trial_value(configured, logical_name)
    stored_value = _constant_trial_value(stored, default_field)
    if configured is not None:
        configured_value = _constant_trial_value(configured, logical_name)
        if not np.isclose(
            stored_value, configured_value, rtol=0.0, atol=1e-15
        ):
            raise MetricError(
                f"{logical_name} disagrees with stored effective configuration"
            )
    return stored_value


def derive_trial_metrics(
    samples: Mapping[str, Any],
    configuration: Mapping[str, Any],
) -> dict[str, Any]:
    """Derive configured endpoint families from one stored trial."""
    enabled = configuration.get("enabled")
    if not isinstance(enabled, list) or not enabled:
        raise MetricError("metric configuration needs a non-empty enabled list")
    result: dict[str, Any] = {"metric_schema_version": METRIC_SCHEMA_VERSION}
    for family in enabled:
        family_config = configuration.get(family, {})
        fields = family_config.get("fields", {})
        if family == "h1":
            thresholds = family_config.get("thresholds", {})
            required_thresholds = (
                "position_threshold_m",
                "orientation_threshold_rad",
                "joint_step_threshold_rad",
                "swivel_step_threshold_rad",
                "input_step_threshold_rad",
            )
            missing = [name for name in required_thresholds if name not in thresholds]
            if missing:
                raise MetricError(f"missing H1 thresholds: {missing}")
            pose_success = _field(
                samples, fields, "pose_success", "map_pose_success"
            )
            differential_valid = _field(
                samples,
                fields,
                "differential_valid",
                "map_differential_valid",
            )
            mapping_valid = np.asarray(pose_success, dtype=bool) & np.asarray(
                differential_valid, dtype=bool
            )
            result.update(
                compute_h1_composite(
                    mapping_valid=mapping_valid,
                    position_error_m=_field(
                        samples,
                        fields,
                        "position_error_m",
                        "map_position_error_m",
                    ),
                    orientation_error_rad=_field(
                        samples,
                        fields,
                        "orientation_error_rad",
                        "map_orientation_error_rad",
                    ),
                    q_slave=_field(samples, fields, "q_slave", "map_q_slave"),
                    swivel_angle_rad=_field(
                        samples,
                        fields,
                        "swivel_angle_rad",
                        "map_swivel_angle_rad",
                    ),
                    master_step_norm=_field(
                        samples,
                        fields,
                        "master_step_norm",
                        "map_master_step_norm",
                    ),
                    accepted=_field(
                        samples,
                        fields,
                        "accepted",
                        "map_accepted",
                        optional=True,
                    ),
                    commanded_reset=_field(
                        samples,
                        fields,
                        "commanded_reset",
                        "map_commanded_reset",
                        optional=True,
                    ),
                    degeneracy_transition=_field(
                        samples,
                        fields,
                        "degeneracy_transition",
                        "map_degeneracy_transition",
                        optional=True,
                    ),
                    **thresholds,
                )
            )
            audit_fields = family_config.get("audit_fields", [])
            if audit_fields:
                required_audit_fields = {
                    "map_validity_reason_code",
                    "map_reach_clip_code",
                    "map_joint_limit_active",
                    "map_geometry_degenerate",
                    "map_slave_min_singular_value",
                    "map_slave_manipulability",
                    "map_low_manipulability",
                }
                if (
                    not isinstance(audit_fields, list)
                    or any(
                        not isinstance(name, str) or not name
                        for name in audit_fields
                    )
                ):
                    raise MetricError("H1 audit_fields must be a list of names")
                configured_audit_fields = set(audit_fields)
                if configured_audit_fields != required_audit_fields:
                    missing_audit = sorted(
                        required_audit_fields - configured_audit_fields
                    )
                    unknown_audit = sorted(
                        configured_audit_fields - required_audit_fields
                    )
                    raise MetricError(
                        "H1 audit_fields must declare the complete audit set; "
                        f"missing={missing_audit}, unknown={unknown_audit}"
                    )
                for stored_name in sorted(required_audit_fields):
                    if stored_name not in samples:
                        raise MetricError(
                            f"missing configured H1 audit field {stored_name!r}"
                        )
                result.update(
                    compute_h1_audit_metrics(
                        differential_valid=differential_valid,
                        validity_reason_code=samples[
                            "map_validity_reason_code"
                        ],
                        reach_clip_code=samples["map_reach_clip_code"],
                        joint_limit_active=samples[
                            "map_joint_limit_active"
                        ],
                        geometry_degenerate=samples[
                            "map_geometry_degenerate"
                        ],
                        low_manipulability=samples[
                            "map_low_manipulability"
                        ],
                        slave_min_singular_value=samples[
                            "map_slave_min_singular_value"
                        ],
                        slave_manipulability=samples[
                            "map_slave_manipulability"
                        ],
                        validity_reason_labels=family_config.get(
                            "validity_reason_labels"
                        ),
                    )
                )
        elif family == "h2":
            result.update(
                compute_h2_wrench_metrics(
                    _field(
                        samples,
                        fields,
                        "wrench_estimated",
                        "wrench_estimated",
                    ),
                    _field(
                        samples,
                        fields,
                        "wrench_reference",
                        "wrench_reference",
                    ),
                    sample_mask=_field(
                        samples,
                        fields,
                        "sample_mask",
                        "wrench_sample_mask",
                        optional=True,
                    ),
                    numerical_rank_deficient=_field(
                        samples,
                        fields,
                        "numerical_rank_deficient",
                        "solver_numerical_rank_deficient",
                        optional=True,
                    ),
                    operationally_ill_conditioned=_field(
                        samples,
                        fields,
                        "operationally_ill_conditioned",
                        "solver_operationally_ill_conditioned",
                        optional=True,
                    ),
                    numerical_rank_threshold=_field(
                        samples,
                        fields,
                        "numerical_rank_threshold",
                        "solver_numerical_rank_threshold",
                        optional=True,
                    ),
                    operational_min_scaled_singular_threshold=_field(
                        samples,
                        fields,
                        "operational_min_scaled_singular_threshold",
                        "operational_min_scaled_singular_threshold",
                        optional=True,
                    ),
                    scaled_singular_values=_field(
                        samples,
                        fields,
                        "scaled_singular_values",
                        "scaled_singular_values",
                        optional=True,
                    ),
                    condition_number=_field(
                        samples,
                        fields,
                        "condition_number",
                        "solver_condition_number",
                        optional=True,
                    ),
                )
            )
        elif family == "h3":
            result.update(
                compute_h3_power_mismatch(
                    tau_master_raw=_field(
                        samples, fields, "tau_master_raw", "tau_master_raw"
                    ),
                    qd_master=_field(samples, fields, "qd_master", "qd_master"),
                    tau_slave_source=_field(
                        samples,
                        fields,
                        "tau_slave_source",
                        "tau_slave_source",
                    ),
                    qd_slave_source=_field(
                        samples,
                        fields,
                        "qd_slave_source",
                        "qd_slave_source",
                    ),
                    dt=_field(samples, fields, "dt", "dt"),
                    return_valid=_field(
                        samples,
                        fields,
                        "return_valid",
                        "return_valid",
                        optional=True,
                    ),
                    force_scale=family_config.get("force_scale", 1.0),
                    epsilon_energy_J=family_config.get(
                        "epsilon_energy_J", 1e-12
                    ),
                    minimum_power_activity_J=family_config.get(
                        "minimum_power_activity_J", 0.0
                    ),
                )
            )
        elif family == "h4":
            energy_min_J = _energy_bound_for_trial(
                samples,
                fields,
                family_config,
                "energy_min_J",
                "configured_energy_min_J",
            )
            energy_max_J = _energy_bound_for_trial(
                samples,
                fields,
                family_config,
                "energy_max_J",
                "configured_energy_max_J",
            )
            result.update(
                audit_h4_energy(
                    energy_before_J=_field(
                        samples, fields, "energy_before_J", "energy_before_J"
                    ),
                    energy_after_J=_field(
                        samples, fields, "energy_after_J", "energy_after_J"
                    ),
                    tau_candidate=_field(
                        samples, fields, "tau_candidate", "tau_master_candidate"
                    ),
                    tau_applied=_field(
                        samples, fields, "tau_applied", "tau_master_applied"
                    ),
                    tau_accepted=_field(
                        samples,
                        fields,
                        "tau_accepted",
                        "tau_master_accepted",
                        optional=True,
                    ),
                    qd_master=_field(samples, fields, "qd_master", "qd_master"),
                    dt=_field(samples, fields, "dt", "dt"),
                    software_preclip_J=_field(
                        samples,
                        fields,
                        "software_preclip_J",
                        "energy_preclip_J",
                        optional=True,
                    ),
                    energy_min_J=energy_min_J,
                    energy_max_J=energy_max_J,
                    epsilon_torque_impulse_Nms=family_config.get(
                        "epsilon_torque_impulse_Nms", 1e-12
                    ),
                    audit_tolerance_J=family_config.get(
                        "audit_tolerance_J", 1e-10
                    ),
                )
            )
        elif family == "bilateral":
            result.update(
                compute_bilateral_diagnostics(
                    master_tracking_error_rad=_field(
                        samples,
                        fields,
                        "master_tracking_error_rad",
                        "master_tracking_error",
                    ),
                    slave_tracking_error_rad=_field(
                        samples,
                        fields,
                        "slave_tracking_error_rad",
                        "slave_tracking_error",
                    ),
                    feedback_torque_Nm=_field(
                        samples,
                        fields,
                        "feedback_torque_Nm",
                        "tau_master_applied",
                    ),
                    contact_force_N=_field(
                        samples,
                        fields,
                        "contact_force_N",
                        "contact_force_norm",
                    ),
                    projection_factor=_field(
                        samples,
                        fields,
                        "projection_factor",
                        "rho",
                    ),
                    projection_tolerance=family_config.get(
                        "projection_tolerance", 1e-12
                    ),
                    contact_force_threshold_N=family_config.get(
                        "contact_force_threshold_N", 1e-6
                    ),
                )
            )
        else:
            raise MetricError(f"unknown metric family {family!r}")
    return result