s23dr_2026/postprocess.py

import numpy as np


def _empty_edges_like(edges: np.ndarray) -> np.ndarray:
    return np.zeros((0, 2), dtype=edges.dtype)


def _compact_vertices_for_edges(verts: np.ndarray, edges: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts[:0], _empty_edges_like(edges)

    used_vertex_indices, inverse = np.unique(edges.reshape(-1), return_inverse=True)
    compact_verts = verts[used_vertex_indices]
    compact_edges = inverse.reshape(-1, 2).astype(edges.dtype, copy=False)
    return compact_verts, compact_edges


def _compact_vertices_and_features_for_edges(
    verts: np.ndarray, edges: np.ndarray, vertex_feats: np.ndarray
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts[:0], _empty_edges_like(edges), vertex_feats[:0]

    used_vertex_indices, inverse = np.unique(edges.reshape(-1), return_inverse=True)
    compact_verts = verts[used_vertex_indices]
    compact_edges = inverse.reshape(-1, 2).astype(edges.dtype, copy=False)
    compact_feats = vertex_feats[used_vertex_indices]
    return compact_verts, compact_edges, compact_feats


def _merged_vertex_from_incident_lines(
    verts: np.ndarray,
    edges: np.ndarray,
    component_members: np.ndarray,
    incident_edge_indices_by_vertex: list[list[int]],
    average_position: np.ndarray,
    intersection_angle_threshold_deg: float | None,
) -> np.ndarray:
    if len(component_members) < 2:
        return average_position
    if intersection_angle_threshold_deg is None or intersection_angle_threshold_deg <= 0:
        return average_position

    component_member_set = set(int(vertex) for vertex in component_members)
    component_edge_indices: set[int] = set()
    for vertex in component_members:
        component_edge_indices.update(incident_edge_indices_by_vertex[int(vertex)])

    line_points: list[np.ndarray] = []
    line_dirs: list[np.ndarray] = []
    for edge_idx in component_edge_indices:
        a, b = edges[edge_idx]
        a = int(a)
        b = int(b)
        if (a in component_member_set) == (b in component_member_set):
            continue

        line_vec = verts[b] - verts[a]
        line_length = np.linalg.norm(line_vec)
        if line_length == 0:
            continue

        line_points.append(verts[a].astype(np.float64, copy=False))
        line_dirs.append(line_vec.astype(np.float64, copy=False) / line_length)

    if len(line_dirs) < 2:
        return average_position

    directions = np.array(line_dirs, dtype=np.float64)
    dot_products = np.abs(directions @ directions.T)
    upper_triangle = np.triu_indices(len(directions), k=1)
    angles_deg = np.degrees(np.arccos(np.clip(dot_products[upper_triangle], -1.0, 1.0)))
    if not np.any(angles_deg >= intersection_angle_threshold_deg):
        return average_position

    system = np.zeros((3, 3), dtype=np.float64)
    rhs = np.zeros(3, dtype=np.float64)
    identity = np.eye(3, dtype=np.float64)
    for point, direction in zip(line_points, line_dirs):
        projection = identity - np.outer(direction, direction)
        system += projection
        rhs += projection @ point

    try:
        intersection = np.linalg.lstsq(system, rhs, rcond=None)[0]
    except np.linalg.LinAlgError:
        return average_position

    if not np.all(np.isfinite(intersection)):
        return average_position
    return intersection.astype(average_position.dtype, copy=False)


def merge_close_vertices(
    verts: np.ndarray,
    edges: np.ndarray,
    distance_threshold: float,
    intersection_angle_threshold_deg: float | None = 30.0,
) -> tuple[np.ndarray, np.ndarray]:
    if len(verts) == 0 or distance_threshold <= 0:
        return verts, edges

    cell_size = float(distance_threshold)
    cell_coords = np.floor(verts / cell_size).astype(np.int64)
    buckets: dict[tuple[int, int, int], list[int]] = {}

    parent = np.arange(len(verts), dtype=np.int64)
    rank = np.zeros(len(verts), dtype=np.int8)
    members = [set([i]) for i in range(len(verts))]
    edge_neighbors = [set() for _ in range(len(verts))]
    incident_edge_indices_by_vertex: list[list[int]] = [[] for _ in range(len(verts))]
    for edge_idx, (a, b) in enumerate(edges):
        a = int(a)
        b = int(b)
        incident_edge_indices_by_vertex[a].append(edge_idx)
        if a != b:
            incident_edge_indices_by_vertex[b].append(edge_idx)
        if a == b:
            continue
        edge_neighbors[a].add(b)
        edge_neighbors[b].add(a)

    def find(i: int) -> int:
        while parent[i] != i:
            parent[i] = parent[parent[i]]
            i = parent[i]
        return i

    def components_share_edge(a_members: set[int], b_members: set[int]) -> bool:
        if len(a_members) > len(b_members):
            a_members, b_members = b_members, a_members
        return any(not edge_neighbors[vertex].isdisjoint(b_members) for vertex in a_members)

    def union(a: int, b: int) -> None:
        ra, rb = find(a), find(b)
        if ra == rb or components_share_edge(members[ra], members[rb]):
            return
        if rank[ra] < rank[rb]:
            ra, rb = rb, ra
        parent[rb] = ra
        if rank[ra] == rank[rb]:
            rank[ra] += 1
        members[ra].update(members[rb])
        members[rb].clear()

    for i, cell in enumerate(cell_coords):
        key = tuple(cell.tolist())
        for dx in (-1, 0, 1):
            for dy in (-1, 0, 1):
                for dz in (-1, 0, 1):
                    neighbor_key = (key[0] + dx, key[1] + dy, key[2] + dz)
                    for j in buckets.get(neighbor_key, ()):
                        if np.linalg.norm(verts[i] - verts[j]) <= distance_threshold:
                            union(i, j)
        buckets.setdefault(key, []).append(i)

    roots = np.array([find(i) for i in range(len(verts))], dtype=np.int64)
    unique_roots, inverse = np.unique(roots, return_inverse=True)
    merged_verts = np.zeros((len(unique_roots), 3), dtype=verts.dtype)
    for new_vertex_idx in range(len(unique_roots)):
        component_members = np.flatnonzero(inverse == new_vertex_idx)
        average_position = np.mean(verts[component_members], axis=0)
        merged_verts[new_vertex_idx] = _merged_vertex_from_incident_lines(
            verts,
            edges,
            component_members,
            incident_edge_indices_by_vertex,
            average_position,
            intersection_angle_threshold_deg,
        )
    old_to_new = inverse
    merged_edges = old_to_new[edges]
    keep = merged_edges[:, 0] != merged_edges[:, 1]
    merged_edges = merged_edges[keep]
    if len(merged_edges) == 0:
        return merged_verts, np.zeros((0, 2), dtype=np.int64)

    merged_edges = np.sort(merged_edges, axis=1)
    merged_edges = np.unique(merged_edges, axis=0)
    return merged_verts, merged_edges


def merge_vertices_with_features(
    verts: np.ndarray,
    edges: np.ndarray,
    vertex_feats: np.ndarray,
    alpha: float = 0.95,
    distance_threshold: float | None = None,
    vertex_scores: np.ndarray | None = None,
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    if len(verts) == 0 or len(vertex_feats) == 0:
        return verts, edges, vertex_feats
    if len(verts) != len(vertex_feats):
        raise ValueError("verts and vertex_feats must have the same length")

    feats = vertex_feats.astype(np.float32, copy=False)
    sim = feats @ feats.T

    parent = np.arange(len(verts), dtype=np.int64)
    rank = np.zeros(len(verts), dtype=np.int8)
    members = [set([i]) for i in range(len(verts))]
    edge_neighbors = [set() for _ in range(len(verts))]
    for a, b in edges:
        if a == b:
            continue
        edge_neighbors[int(a)].add(int(b))
        edge_neighbors[int(b)].add(int(a))

    def find(i: int) -> int:
        while parent[i] != i:
            parent[i] = parent[parent[i]]
            i = parent[i]
        return i

    def components_share_edge(a_members: set[int], b_members: set[int]) -> bool:
        if len(a_members) > len(b_members):
            a_members, b_members = b_members, a_members
        return any(not edge_neighbors[vertex].isdisjoint(b_members) for vertex in a_members)

    def union(a: int, b: int) -> None:
        ra, rb = find(a), find(b)
        if ra == rb or components_share_edge(members[ra], members[rb]):
            return
        if rank[ra] < rank[rb]:
            ra, rb = rb, ra
        parent[rb] = ra
        if rank[ra] == rank[rb]:
            rank[ra] += 1
        members[ra].update(members[rb])
        members[rb].clear()

    rows, cols = np.where(np.triu(sim >= alpha, k=1))
    if distance_threshold is not None:
        if distance_threshold <= 0:
            rows, cols = np.array([], dtype=np.int64), np.array([], dtype=np.int64)
        else:
            distances = np.linalg.norm(verts[rows] - verts[cols], axis=1)
            keep = distances <= distance_threshold
            rows, cols = rows[keep], cols[keep]
    for a, b in zip(rows.tolist(), cols.tolist()):
        union(a, b)

    roots = np.array([find(i) for i in range(len(verts))], dtype=np.int64)
    unique_roots, inverse = np.unique(roots, return_inverse=True)
    merged_verts = np.zeros((len(unique_roots), 3), dtype=verts.dtype)
    counts = np.bincount(inverse)
    np.add.at(merged_verts, inverse, verts)
    merged_verts /= counts[:, None]

    if vertex_scores is not None:
        scores = vertex_scores.astype(np.float32, copy=False)
        best_score = np.full(len(unique_roots), -np.inf, dtype=np.float32)
        best_index = np.zeros(len(unique_roots), dtype=np.int64)
        for idx, group_idx in enumerate(inverse):
            score = scores[idx]
            if score > best_score[group_idx]:
                best_score[group_idx] = score
                best_index[group_idx] = idx
        merged_feats = vertex_feats[best_index]
    else:
        merged_feats = np.zeros((len(unique_roots), vertex_feats.shape[1]), dtype=vertex_feats.dtype)
        np.add.at(merged_feats, inverse, vertex_feats)
        merged_feats /= counts[:, None]

    merged_edges = inverse[edges]
    keep = merged_edges[:, 0] != merged_edges[:, 1]
    merged_edges = merged_edges[keep]
    if len(merged_edges) == 0:
        return merged_verts, np.zeros((0, 2), dtype=np.int64), merged_feats

    merged_edges = np.sort(merged_edges, axis=1)
    merged_edges = np.unique(merged_edges, axis=0)
    return merged_verts, merged_edges, merged_feats


def filter_edges_by_length(
    verts: np.ndarray,
    edges: np.ndarray,
    min_length: float | None = 0.05,
    max_length: float | None = 100.0,
) -> tuple[np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts[:0], _empty_edges_like(edges)

    edge_segments = verts[edges]
    edge_lengths = np.linalg.norm(edge_segments[:, 1] - edge_segments[:, 0], axis=1)
    keep = np.ones(len(edges), dtype=bool)
    if min_length is not None:
        keep &= edge_lengths >= min_length
    if max_length is not None:
        keep &= edge_lengths <= max_length
    return _compact_vertices_for_edges(verts, edges[keep])


def add_gap_fill_edges_from_candidates(
    verts: np.ndarray,
    edges: np.ndarray,
    candidate_edge_points: np.ndarray,
    merge_distance_threshold: float,
) -> tuple[np.ndarray, np.ndarray]:
    if len(verts) == 0 or len(candidate_edge_points) == 0 or merge_distance_threshold < 0:
        return verts, edges

    candidate_edge_points = np.asarray(candidate_edge_points).reshape(-1, 2, verts.shape[1])
    endpoint_distances = np.linalg.norm(
        candidate_edge_points.reshape(-1, verts.shape[1])[:, None, :] - verts[None, :, :],
        axis=2,
    )
    nearest_vertices = np.argmin(endpoint_distances, axis=1).reshape(-1, 2)
    nearest_distances = np.min(endpoint_distances, axis=1).reshape(-1, 2)

    known_edges = {tuple(sorted((int(a), int(b)))) for a, b in edges if int(a) != int(b)}
    added_edges: list[tuple[int, int]] = []
    for edge_vertices, edge_distances in zip(nearest_vertices, nearest_distances):
        if np.any(edge_distances > merge_distance_threshold):
            continue

        a, b = int(edge_vertices[0]), int(edge_vertices[1])
        if a == b:
            continue

        edge_key = tuple(sorted((a, b)))
        if edge_key in known_edges:
            continue

        known_edges.add(edge_key)
        added_edges.append(edge_key)

    if len(added_edges) == 0:
        return verts, edges

    added_edges_array = np.array(added_edges, dtype=edges.dtype)
    return verts, np.concatenate([edges, added_edges_array], axis=0)


def add_edges_by_translated_edge_symmetry(
    verts: np.ndarray,
    edges: np.ndarray,
    distance_threshold: float = 0.2,
) -> tuple[np.ndarray, np.ndarray]:
    if len(verts) == 0 or len(edges) == 0 or distance_threshold < 0:
        return verts, edges

    template_edges = edges.copy()
    known_edges = {tuple(sorted((int(a), int(b)))) for a, b in edges if int(a) != int(b)}
    added_edges: list[tuple[int, int]] = []
    for a, b in template_edges:
        a = int(a)
        b = int(b)
        edge_vector = verts[b] - verts[a]
        if not np.any(edge_vector):
            continue

        for template_vector in (edge_vector, -edge_vector):
            translated_endpoints = verts + template_vector
            distances = np.linalg.norm(translated_endpoints[:, None, :] - verts[None, :, :], axis=2)
            nearest_vertices = np.argmin(distances, axis=1)
            nearest_distances = np.min(distances, axis=1)
            for start_vertex, (end_vertex, end_distance) in enumerate(zip(nearest_vertices, nearest_distances)):
                if end_distance > distance_threshold:
                    continue

                end_vertex = int(end_vertex)
                if start_vertex == end_vertex:
                    continue

                edge_key = tuple(sorted((start_vertex, end_vertex)))
                if edge_key in known_edges:
                    continue

                known_edges.add(edge_key)
                added_edges.append(edge_key)

    if len(added_edges) == 0:
        return verts, edges

    added_edges_array = np.array(added_edges, dtype=edges.dtype)
    return verts, np.concatenate([edges, added_edges_array], axis=0)


def _edge_angle_to_xz_plane_deg(edge_vecs: np.ndarray) -> np.ndarray:
    xz_norms = np.linalg.norm(edge_vecs[:, [0, 2]], axis=1)
    return np.degrees(np.arctan2(np.abs(edge_vecs[:, 1]), xz_norms))


def _infer_principal_xz_directions(
    edge_vecs: np.ndarray,
    edge_lengths: np.ndarray,
    horizontal_angle_threshold_deg: float,
) -> tuple[np.ndarray, np.ndarray] | None:
    xz_vecs = edge_vecs[:, [0, 2]]
    xz_norms = np.linalg.norm(xz_vecs, axis=1)
    angles = _edge_angle_to_xz_plane_deg(edge_vecs)
    keep = (edge_lengths > 0) & (xz_norms > 0) & (angles <= horizontal_angle_threshold_deg)
    if not np.any(keep):
        return None

    dirs = xz_vecs[keep] / xz_norms[keep, None]
    weights = edge_lengths[keep]
    covariance = np.einsum("i,ij,ik->jk", weights, dirs, dirs)
    if not np.any(covariance):
        return None

    eigvals, eigvecs = np.linalg.eigh(covariance)
    primary = eigvecs[:, int(np.argmax(eigvals))]
    primary_norm = np.linalg.norm(primary)
    if primary_norm == 0:
        return None

    primary = primary / primary_norm
    secondary = np.array([-primary[1], primary[0]], dtype=primary.dtype)
    return primary, secondary


def project_near_horizontal_edges_to_xz_plane(
    verts: np.ndarray,
    edges: np.ndarray,
    horizontal_angle_threshold_deg: float = 5.0,
    principal_direction_angle_threshold_deg: float = 5.0,
) -> tuple[np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts, edges

    edge_vecs = verts[edges[:, 1]] - verts[edges[:, 0]]
    edge_lengths = np.linalg.norm(edge_vecs, axis=1)
    principal_dirs = _infer_principal_xz_directions(edge_vecs, edge_lengths, horizontal_angle_threshold_deg)
    if principal_dirs is None:
        return verts, edges

    xz_vecs = edge_vecs[:, [0, 2]]
    xz_norms = np.linalg.norm(xz_vecs, axis=1)
    valid = (edge_lengths > 0) & (xz_norms > 0)
    if not np.any(valid):
        return verts, edges

    angles_to_xz = _edge_angle_to_xz_plane_deg(edge_vecs)
    dirs_xz = np.zeros_like(xz_vecs, dtype=np.result_type(verts.dtype, np.float64))
    dirs_xz[valid] = xz_vecs[valid] / xz_norms[valid, None]
    dots = np.stack([np.abs(dirs_xz @ direction) for direction in principal_dirs], axis=1)
    angles_to_principal = np.degrees(np.arccos(np.clip(np.max(dots, axis=1), -1.0, 1.0)))
    eligible = (
        valid
        & (angles_to_xz <= horizontal_angle_threshold_deg)
        & (angles_to_principal <= principal_direction_angle_threshold_deg)
    )
    eligible_edges = edges[eligible]
    if len(eligible_edges) == 0:
        return verts, edges

    parent = np.arange(len(verts), dtype=np.int64)

    def find(i: int) -> int:
        while parent[i] != i:
            parent[i] = parent[parent[i]]
            i = parent[i]
        return i

    def union(a: int, b: int) -> None:
        ra, rb = find(a), find(b)
        if ra != rb:
            parent[rb] = ra

    for a, b in eligible_edges:
        union(int(a), int(b))

    new_verts = verts.copy()
    eligible_vertices = np.unique(eligible_edges.reshape(-1))
    roots = np.array([find(int(vertex)) for vertex in eligible_vertices], dtype=np.int64)
    for root in np.unique(roots):
        component_vertices = eligible_vertices[roots == root]
        new_verts[component_vertices, 1] = np.mean(verts[component_vertices, 1])
    return new_verts, edges


def snap_edges_to_principal_xz_directions(
    verts: np.ndarray,
    edges: np.ndarray,
    horizontal_angle_threshold_deg: float = 5.0,
    principal_direction_angle_threshold_deg: float = 5.0,
) -> tuple[np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts, edges

    edge_vecs = verts[edges[:, 1]] - verts[edges[:, 0]]
    edge_lengths = np.linalg.norm(edge_vecs, axis=1)
    principal_dirs = _infer_principal_xz_directions(edge_vecs, edge_lengths, horizontal_angle_threshold_deg)
    if principal_dirs is None:
        return verts, edges

    xz_vecs = edge_vecs[:, [0, 2]]
    xz_norms = np.linalg.norm(xz_vecs, axis=1)
    valid = xz_norms > 0
    if not np.any(valid):
        return verts, edges

    dirs_xz = np.zeros_like(xz_vecs, dtype=np.result_type(verts.dtype, np.float64))
    dirs_xz[valid] = xz_vecs[valid] / xz_norms[valid, None]
    dots = np.stack([np.abs(dirs_xz @ direction) for direction in principal_dirs], axis=1)
    best_direction_indices = np.argmax(dots, axis=1)
    angles_to_principal = np.degrees(np.arccos(np.clip(np.max(dots, axis=1), -1.0, 1.0)))
    eligible = valid & (angles_to_principal <= principal_direction_angle_threshold_deg)
    eligible_edges = edges[eligible]
    if len(eligible_edges) == 0:
        return verts, edges

    eligible_vertices = np.unique(eligible_edges.reshape(-1))
    vertex_to_local = {int(vertex): idx for idx, vertex in enumerate(eligible_vertices)}
    constraints = np.zeros((len(eligible_edges), 2 * len(eligible_vertices)), dtype=np.float64)
    eligible_direction_indices = best_direction_indices[eligible]
    for row, ((a, b), direction_idx) in enumerate(zip(eligible_edges, eligible_direction_indices)):
        direction = principal_dirs[int(direction_idx)]
        normal = np.array([-direction[1], direction[0]], dtype=np.float64)
        a_col = 2 * vertex_to_local[int(a)]
        b_col = 2 * vertex_to_local[int(b)]
        constraints[row, a_col : a_col + 2] = -normal
        constraints[row, b_col : b_col + 2] = normal

    original_xz = verts[eligible_vertices][:, [0, 2]].astype(np.float64, copy=False).reshape(-1)
    residual = constraints @ original_xz
    if not np.any(residual):
        return verts, edges

    system = constraints @ constraints.T
    correction_weights = np.linalg.lstsq(system, residual, rcond=None)[0]
    snapped_xz = original_xz - constraints.T @ correction_weights

    new_verts = verts.copy()
    snapped_xz = snapped_xz.reshape(-1, 2)
    new_verts[eligible_vertices, 0] = snapped_xz[:, 0]
    new_verts[eligible_vertices, 2] = snapped_xz[:, 1]
    return new_verts, edges


project_near_horizontal_edges_to_xy_plane = project_near_horizontal_edges_to_xz_plane
snap_edges_to_principal_xy_directions = snap_edges_to_principal_xz_directions


def _remove_solitary_edges(verts: np.ndarray, edges: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
    if len(edges) == 0:
        return verts[:0], _empty_edges_like(edges)

    degrees = np.bincount(edges.reshape(-1), minlength=len(verts))
    edge_degrees = degrees[edges]
    keep = np.any(edge_degrees > 1, axis=1)
    return _compact_vertices_for_edges(verts, edges[keep])


def remove_solitary_edges(verts: np.ndarray, edges: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
    return _remove_solitary_edges(verts, edges)


def filter_edges_by_connectivity(
    verts: np.ndarray,
    edges: np.ndarray,
    min_endpoint_degree: int = 2,
) -> tuple[np.ndarray, np.ndarray]:
    """Keep edges where at least one endpoint connects to >= min_endpoint_degree other edges.

    Boosts structurally connected edges; penalizes isolated ones. min_endpoint_degree=2
    keeps any edge touching a vertex with 2+ incident edges.
    """
    if len(edges) == 0 or min_endpoint_degree <= 1:
        return verts, edges

    degrees = np.bincount(edges.reshape(-1), minlength=len(verts))
    endpoint_degrees = degrees[edges]  # [E, 2]
    keep = np.any(endpoint_degrees >= min_endpoint_degree, axis=1)
    return _compact_vertices_for_edges(verts, edges[keep])


def _point_to_segments_distance(points: np.ndarray, seg_starts: np.ndarray, seg_ends: np.ndarray) -> np.ndarray:
    dtype = np.result_type(points.dtype, seg_starts.dtype, seg_ends.dtype, np.float64)
    points = points.astype(dtype, copy=False)
    seg_starts = seg_starts.astype(dtype, copy=False)
    seg_ends = seg_ends.astype(dtype, copy=False)

    seg_vecs = seg_ends - seg_starts
    seg_lens_sq = np.einsum("ij,ij->i", seg_vecs, seg_vecs)
    offsets = points[:, None, :] - seg_starts[None, :, :]

    t = np.zeros((len(points), len(seg_starts)), dtype=dtype)
    valid = seg_lens_sq > 0
    if np.any(valid):
        t[:, valid] = np.einsum("pkj,kj->pk", offsets[:, valid], seg_vecs[valid]) / seg_lens_sq[valid]
    t = np.clip(t, 0.0, 1.0)

    closest = seg_starts[None, :, :] + t[:, :, None] * seg_vecs[None, :, :]
    return np.linalg.norm(points[:, None, :] - closest, axis=2)


def remove_edges_close_to_other_edges(
    verts: np.ndarray, edges: np.ndarray, distance_threshold: float
) -> tuple[np.ndarray, np.ndarray]:
    if distance_threshold <= 0:
        return verts, edges
    if len(edges) == 0:
        return verts[:0], _empty_edges_like(edges)

    edge_segments = verts[edges]
    edge_vecs = edge_segments[:, 1] - edge_segments[:, 0]
    edge_lengths = np.linalg.norm(edge_vecs, axis=1)
    edge_order = np.lexsort((np.arange(len(edges)), -edge_lengths))

    kept_edge_indices: list[int] = []
    for edge_idx in edge_order:
        if len(kept_edge_indices) == 0:
            kept_edge_indices.append(edge_idx)
            continue

        endpoints = edge_segments[edge_idx]
        kept_segments = edge_segments[kept_edge_indices]
        distances = _point_to_segments_distance(endpoints, kept_segments[:, 0], kept_segments[:, 1])
        if np.any(distances.mean(axis=0) <= distance_threshold):
            continue
        kept_edge_indices.append(edge_idx)

    kept_edges = edges[sorted(kept_edge_indices)]
    return _compact_vertices_for_edges(verts, kept_edges)


def post_process_wireframe(
    verts: np.ndarray,
    edges: np.ndarray,
    merge_distance_threshold: float,
    remove_solitary_edges: bool = False,
    filter_by_length: bool = False,
    project_near_horizontal_edges: bool = False,
    snap_to_principal_directions: bool = False,
    min_edge_length: float | None = 0.05,
    max_edge_length: float | None = 100.0,
    horizontal_angle_threshold_deg: float = 5.0,
    principal_direction_angle_threshold_deg: float = 5.0,
    intersection_angle_threshold_deg: float | None = 30.0,
    gap_fill_candidate_edge_points: np.ndarray | None = None,
    complete_symmetric_edges: bool = False,
    symmetric_edge_distance_threshold: float = 0.2,
    connectivity_filter_min_degree: int = 0,
) -> tuple[np.ndarray, np.ndarray]:
    verts, edges = merge_close_vertices(
        verts,
        edges,
        merge_distance_threshold,
        intersection_angle_threshold_deg=intersection_angle_threshold_deg,
    )
    if complete_symmetric_edges:
        verts, edges = add_edges_by_translated_edge_symmetry(
            verts,
            edges,
            distance_threshold=symmetric_edge_distance_threshold,
        )
    if filter_by_length:
        verts, edges = filter_edges_by_length(verts, edges, min_edge_length, max_edge_length)
    if project_near_horizontal_edges:
        verts, edges = project_near_horizontal_edges_to_xz_plane(
            verts,
            edges,
            horizontal_angle_threshold_deg=horizontal_angle_threshold_deg,
            principal_direction_angle_threshold_deg=principal_direction_angle_threshold_deg,
        )
    if snap_to_principal_directions:
        verts, edges = snap_edges_to_principal_xz_directions(
            verts,
            edges,
            horizontal_angle_threshold_deg=horizontal_angle_threshold_deg,
            principal_direction_angle_threshold_deg=principal_direction_angle_threshold_deg,
        )
    if remove_solitary_edges:
        verts, edges = _remove_solitary_edges(verts, edges)
    if gap_fill_candidate_edge_points is not None:
        verts, edges = add_gap_fill_edges_from_candidates(
            verts,
            edges,
            gap_fill_candidate_edge_points,
            merge_distance_threshold,
        )
    if connectivity_filter_min_degree > 1:
        verts, edges = filter_edges_by_connectivity(verts, edges, min_endpoint_degree=connectivity_filter_min_degree)
    return verts, edges