CAD / CODE /evaluation /eval /eval_geometric04.py

Add frozen CODE evaluation toolkit and portable full-GT scoring CLI

15d1f1d verified 4 days ago

25 kB

#!/usr/bin/env python3
"""Representation-invariant metric helpers for the direct ezdxf task.
Metric v4 is deliberately versioned beside v2/v3.  Its first validated scope is
the DISK family.  The important change is that a mathematically straight edge
is scored as one edge regardless of how many contiguous DXF LINE entities were
used to encode it.  Rich DIMENSION extraction also lets the evaluator verify
that displayed values agree with measured drawing geometry under one sheet
scale, instead of checking dimension text alone.
"""

from __future__ import annotations

import math
from collections import defaultdict
from pathlib import Path
from typing import Any, Iterable

import ezdxf
import numpy as np

try:  # package import used by tests
    from . import eval_geometric02 as EV2
    from . import eval_geometric03 as EV3
except ImportError:  # direct/module import used by reward workers
    import eval_geometric02 as EV2
    import eval_geometric03 as EV3


CANONICAL_ABS_TOL = 1e-6
CANONICAL_REL_TOL = 1e-6
DIMENSION_SCALE_INLIER_REL = 0.05
DIMENSION_SCALE_DECAY = 0.05
DISK_RATIO_DECAY = 0.12


def _visibility(layer: Any) -> str:
    return "hidden" if "hidden" in str(layer).lower() else "visible"


def _clean(value: float) -> float:
    value = round(float(value), 12)
    return 0.0 if abs(value) < 5e-13 else value


def _empty_entities() -> dict[str, list[Any]]:
    return {"lines": [], "circles": [], "arcs": [], "dims": []}


def extract_entities(dxf_path: str | Path) -> dict[str, Any] | None:
    """Extract primitives plus lossless-enough DIMENSION diagnostics.
    Unlike v2, every dimension is a dictionary containing its raw ezdxf
    measurement and all available definition/reference points.  Measurements
    are paper/model-space lengths; displayed text is usually a physical length,
    so their ratio (not equality) is checked by :func:`score_dimension_consistency`.
    """

    try:
        doc = ezdxf.readfile(str(dxf_path))
    except Exception:
        return None
    entities = _empty_entities()
    for entity in doc.modelspace():
        entity_type = entity.dxftype()
        layer = entity.dxf.layer if entity.dxf.hasattr("layer") else "0"
        if entity_type == "LINE":
            entities["lines"].append(
                (
                    float(entity.dxf.start.x),
                    float(entity.dxf.start.y),
                    float(entity.dxf.end.x),
                    float(entity.dxf.end.y),
                    layer,
                )
            )
        elif entity_type == "CIRCLE":
            entities["circles"].append(
                (
                    float(entity.dxf.center.x),
                    float(entity.dxf.center.y),
                    float(entity.dxf.radius),
                    layer,
                )
            )
        elif entity_type == "ARC":
            entities["arcs"].append(
                (
                    float(entity.dxf.center.x),
                    float(entity.dxf.center.y),
                    float(entity.dxf.radius),
                    float(entity.dxf.start_angle),
                    float(entity.dxf.end_angle),
                    layer,
                )
            )
        elif entity_type == "DIMENSION":
            dimtype = int(entity.dxf.dimtype) if entity.dxf.hasattr("dimtype") else 0
            text = str(entity.dxf.text) if entity.dxf.hasattr("text") else ""
            points: dict[str, tuple[float, float]] = {}
            for name in (
                "defpoint",
                "defpoint2",
                "defpoint3",
                "defpoint4",
                "defpoint5",
            ):
                if entity.dxf.hasattr(name):
                    point = getattr(entity.dxf, name)
                    points[name] = (float(point.x), float(point.y))
            base = dimtype & 0x0F
            if base in (0, 1) and "defpoint2" in points and "defpoint3" in points:
                first, second = points["defpoint2"], points["defpoint3"]
                binding = ((first[0] + second[0]) / 2, (first[1] + second[1]) / 2)
            elif base in (3, 4) and "defpoint4" in points:
                binding = points["defpoint4"]
            else:
                binding = points.get("defpoint", (0.0, 0.0))
            try:
                measurement = float(entity.get_measurement())
                if not math.isfinite(measurement):
                    measurement = None
            except Exception:
                measurement = None
            text_family, display_value, _ = EV2.parse_dim_value(text)
            type_family = EV2.dim_type_name(dimtype)
            family = (
                type_family
                if type_family in {"linear", "aligned", "diameter", "radius"}
                else text_family
            )
            entities["dims"].append(
                {
                    "dimtype": dimtype,
                    "family": family,
                    "text": text,
                    "display_value": display_value,
                    "measurement": measurement,
                    "layer": layer,
                    "definition_point": points.get("defpoint", (0.0, 0.0)),
                    "binding_point": binding,
                    "ref_points": points,
                }
            )
    return entities


def canonicalize_entities(
    entities: dict[str, Any],
    *,
    abs_tol: float = CANONICAL_ABS_TOL,
    rel_tol: float = CANONICAL_REL_TOL,
) -> dict[str, Any]:
    """Merge overlapping/touching collinear LINEs without bridging real gaps.
    Visible and hidden geometry are isolated, while arbitrary visible layer
    names are treated as representation detail.  The operation is deterministic
    and idempotent.  Circles, arcs and rich dimensions are retained unchanged.
    """

    result = {
        "lines": [],
        "circles": list(entities.get("circles", [])),
        "arcs": list(entities.get("arcs", [])),
        "dims": list(entities.get("dims", [])),
    }
    diagonal = EV2.bbox_diagonal(
        {
            "lines": list(entities.get("lines", [])),
            "circles": result["circles"],
            "arcs": result["arcs"],
            "dims": [],
        }
    )
    tolerance = max(float(abs_tol), float(rel_tol) * max(diagonal, 1.0))
    angular_tolerance = max(1e-10, tolerance / max(diagonal, 1.0))

    clusters: list[dict[str, Any]] = []
    ordered_lines = sorted(
        entities.get("lines", []),
        key=lambda line: (
            _visibility(line[4]),
            min(float(line[0]), float(line[2])),
            min(float(line[1]), float(line[3])),
            max(float(line[0]), float(line[2])),
            max(float(line[1]), float(line[3])),
            str(line[4]),
        ),
    )
    for x1, y1, x2, y2, layer in ordered_lines:
        first = np.asarray([float(x1), float(y1)], dtype=float)
        second = np.asarray([float(x2), float(y2)], dtype=float)
        delta = second - first
        length = float(np.linalg.norm(delta))
        if not math.isfinite(length) or length <= 1e-12:
            continue
        direction = delta / length
        if direction[0] < -angular_tolerance or (
            abs(direction[0]) <= angular_tolerance and direction[1] < 0
        ):
            direction = -direction
        normal = np.asarray([-direction[1], direction[0]], dtype=float)
        offset = float(np.dot(normal, first))
        visibility = _visibility(layer)
        selected = None
        for cluster in clusters:
            if cluster["visibility"] != visibility:
                continue
            cross = abs(float(np.cross(cluster["direction"], direction)))
            if cross > angular_tolerance:
                continue
            # Recompute offset using the cluster normal so reversed/nearly-equal
            # directions cannot split an otherwise identical support line.
            cluster_offset = float(np.dot(cluster["normal"], first))
            if abs(cluster_offset - cluster["offset"]) <= tolerance:
                selected = cluster
                break
        if selected is None:
            selected = {
                "visibility": visibility,
                "direction": direction,
                "normal": normal,
                "offset": offset,
                "intervals": [],
                "layers": [],
            }
            clusters.append(selected)
        start = float(np.dot(selected["direction"], first))
        end = float(np.dot(selected["direction"], second))
        selected["intervals"].append((min(start, end), max(start, end)))
        selected["layers"].append(str(layer))

    merged: list[tuple[float, float, float, float, str]] = []
    for cluster in clusters:
        intervals = sorted(cluster["intervals"])
        if not intervals:
            continue
        union: list[list[float]] = [[intervals[0][0], intervals[0][1]]]
        for start, end in intervals[1:]:
            if start <= union[-1][1] + tolerance:
                union[-1][1] = max(union[-1][1], end)
            else:
                union.append([start, end])
        output_layer = "HIDDEN" if cluster["visibility"] == "hidden" else min(cluster["layers"])
        for start, end in union:
            first = cluster["direction"] * start + cluster["normal"] * cluster["offset"]
            second = cluster["direction"] * end + cluster["normal"] * cluster["offset"]
            merged.append(
                (
                    _clean(first[0]),
                    _clean(first[1]),
                    _clean(second[0]),
                    _clean(second[1]),
                    output_layer,
                )
            )
    result["lines"] = sorted(
        merged,
        key=lambda line: (_visibility(line[4]), *line[:4], str(line[4])),
    )
    return result


def _point_segment_distance(
    point: tuple[float, float], line: tuple[Any, ...]
) -> float:
    value = np.asarray(point, dtype=float)
    first = np.asarray(line[:2], dtype=float)
    second = np.asarray(line[2:4], dtype=float)
    delta = second - first
    denominator = float(np.dot(delta, delta))
    if denominator <= 1e-18:
        return float(np.linalg.norm(value - first))
    fraction = float(np.dot(value - first, delta) / denominator)
    projection = first + min(1.0, max(0.0, fraction)) * delta
    return float(np.linalg.norm(value - projection))


def _distance_to_geometry(point: tuple[float, float], entities: dict[str, Any]) -> float:
    distances: list[float] = []
    for line in entities.get("lines", []):
        if _visibility(line[-1]) == "visible":
            distances.append(_point_segment_distance(point, line))
    for cx, cy, radius, layer in entities.get("circles", []):
        if _visibility(layer) == "visible":
            radial = math.dist(point, (float(cx), float(cy)))
            distances.append(abs(radial - abs(float(radius))))
            distances.append(radial)  # centre references are valid for circular dims
    for cx, cy, radius, start, end, layer in entities.get("arcs", []):
        if _visibility(layer) == "visible":
            distances.append(abs(math.dist(point, (float(cx), float(cy))) - abs(float(radius))))
    return min(distances) if distances else math.inf


def _dimension_reference_points(dimension: dict[str, Any]) -> list[tuple[float, float]]:
    refs = dimension.get("ref_points") or {}
    family = str(dimension.get("family", ""))
    if family in {"linear", "aligned"}:
        return [refs[name] for name in ("defpoint2", "defpoint3") if name in refs]
    if family in {"diameter", "radius"}:
        points = [refs[name] for name in ("defpoint4", "defpoint") if name in refs]
        return points or list(refs.values())
    return list(refs.values())


def score_dimension_consistency(entities: dict[str, Any]) -> dict[str, Any]:
    """Check shared drawing scale and whether dimensions bind to the outline."""

    dimensions = entities.get("dims", [])
    if not dimensions:
        return {
            "score": 1.0,
            "n_dimensions": 0,
            "n_valid": 0,
            "n_invalid": 0,
            "drawing_scale": None,
            "scale_consistency": 1.0,
            "binding_quality": 1.0,
            "data_valid": True,
            "items": [],
        }
    diagonal = max(EV2.bbox_diagonal(entities), 1.0)
    binding_sigma = max(1e-6, 0.005 * diagonal)
    items: list[dict[str, Any]] = []
    valid_logs: list[float] = []
    for dimension in dimensions:
        value = dimension.get("display_value")
        measurement = dimension.get("measurement")
        family = str(dimension.get("family", ""))
        valid = (
            family in {"linear", "aligned", "diameter", "radius"}
            and isinstance(value, (int, float))
            and isinstance(measurement, (int, float))
            and math.isfinite(float(value))
            and math.isfinite(float(measurement))
            and abs(float(value)) > 1e-9
            and abs(float(measurement)) > 1e-9
        )
        ratio = abs(float(measurement) / float(value)) if valid else None
        if ratio is not None and ratio > 0:
            valid_logs.append(math.log(ratio))
        references = _dimension_reference_points(dimension)
        reference_distances = [_distance_to_geometry(point, entities) for point in references]
        finite_distances = [distance for distance in reference_distances if math.isfinite(distance)]
        if finite_distances:
            # Every extension/reference endpoint should be attached.  A mean
            # prevents one valid endpoint from hiding a detached second endpoint.
            binding = float(np.mean(np.exp(-np.asarray(finite_distances) / binding_sigma)))
        else:
            binding = 0.0
        items.append(
            {
                "family": family,
                "text": dimension.get("text", ""),
                "display_value": value,
                "measurement": measurement,
                "scale_ratio": ratio,
                "valid": bool(valid),
                "binding_quality": round(binding, 6),
                "reference_distance_mean": (
                    round(float(np.mean(finite_distances)), 6) if finite_distances else None
                ),
            }
        )

    if valid_logs:
        threshold = math.log1p(DIMENSION_SCALE_INLIER_REL)
        candidates: list[tuple[int, float, list[float]]] = []
        global_median = float(np.median(valid_logs))
        for candidate in valid_logs:
            inliers = [value for value in valid_logs if abs(value - candidate) <= threshold]
            centre = float(np.median(inliers))
            candidates.append((len(inliers), -abs(centre - global_median), inliers))
        _, _, consensus = max(candidates, key=lambda value: (value[0], value[1]))
        centre_log = float(np.median(consensus))
        qualities = [
            math.exp(-abs(value - centre_log) / DIMENSION_SCALE_DECAY)
            for value in valid_logs
        ]
        scale_consistency = float(np.mean(qualities))
        drawing_scale = math.exp(centre_log)
    else:
        centre_log = None
        scale_consistency = 0.0
        drawing_scale = None
    valid_index = 0
    for item in items:
        if item["scale_ratio"] is None or centre_log is None:
            item["scale_inlier"] = False
            item["scale_quality"] = 0.0
            continue
        log_ratio = valid_logs[valid_index]
        valid_index += 1
        item["scale_inlier"] = abs(log_ratio - centre_log) <= math.log1p(DIMENSION_SCALE_INLIER_REL)
        item["scale_quality"] = round(
            math.exp(-abs(log_ratio - centre_log) / DIMENSION_SCALE_DECAY), 6
        )
    binding_quality = float(np.mean([item["binding_quality"] for item in items]))
    n_valid = len(valid_logs)
    n_invalid = len(dimensions) - n_valid
    completeness = n_valid / max(len(dimensions), 1)
    score = scale_consistency * binding_quality * completeness
    return {
        "score": round(float(score), 6),
        "n_dimensions": len(dimensions),
        "n_valid": n_valid,
        "n_invalid": n_invalid,
        "drawing_scale": round(float(drawing_scale), 8) if drawing_scale is not None else None,
        "scale_consistency": round(float(scale_consistency), 6),
        "binding_quality": round(float(binding_quality), 6),
        "data_valid": bool(n_invalid == 0 and scale_consistency >= 0.8),
        "items": items,
    }


def _connected_line_components(lines: list[tuple[Any, ...]], tolerance: float) -> list[list[tuple[Any, ...]]]:
    if not lines:
        return []
    parent = list(range(len(lines)))

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

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

    endpoints = [
        (np.asarray(line[:2], dtype=float), np.asarray(line[2:4], dtype=float))
        for line in lines
    ]
    for first in range(len(lines)):
        for second in range(first + 1, len(lines)):
            if any(
                float(np.linalg.norm(a - b)) <= tolerance
                for a in endpoints[first]
                for b in endpoints[second]
            ):
                union(first, second)
    groups: dict[int, list[tuple[Any, ...]]] = defaultdict(list)
    for index, line in enumerate(lines):
        groups[find(index)].append(line)
    return list(groups.values())


def _rectangle_sides(entities: dict[str, Any]) -> list[tuple[float, float]]:
    visible = [line for line in entities.get("lines", []) if _visibility(line[-1]) == "visible"]
    tolerance = max(1e-6, 1e-5 * max(EV2.bbox_diagonal(entities), 1.0))
    rectangles: list[tuple[float, float]] = []
    for component in _connected_line_components(visible, tolerance):
        points = np.asarray([point for line in component for point in (line[:2], line[2:4])], dtype=float)
        low = points.min(axis=0)
        high = points.max(axis=0)
        width, height = float(high[0] - low[0]), float(high[1] - low[1])
        if width <= tolerance or height <= tolerance:
            continue
        horizontal = vertical = 0.0
        for x1, y1, x2, y2, _ in component:
            if abs(float(y2) - float(y1)) <= tolerance:
                horizontal += abs(float(x2) - float(x1))
            elif abs(float(x2) - float(x1)) <= tolerance:
                vertical += abs(float(y2) - float(y1))
        if horizontal >= 1.8 * width and vertical >= 1.8 * height:
            rectangles.append((width, height))
    return rectangles


def _count_f1(predicted: int, target: int) -> float:
    if predicted == 0 and target == 0:
        return 1.0
    if predicted == 0 or target == 0:
        return 0.0
    precision = min(predicted, target) / predicted
    recall = min(predicted, target) / target
    return 2 * precision * recall / (precision + recall)


def _disk_features(entities: dict[str, Any]) -> dict[str, Any]:
    circles = [circle for circle in entities.get("circles", []) if _visibility(circle[-1]) == "visible"]
    rectangles = _rectangle_sides(entities)
    circle_diameter = float(np.median([2 * abs(float(circle[2])) for circle in circles])) if circles else None
    diameter_sides: list[float] = []
    thickness_sides: list[float] = []
    if circle_diameter and circle_diameter > 1e-9:
        for width, height in rectangles:
            if abs(math.log(max(width, 1e-12) / circle_diameter)) <= abs(
                math.log(max(height, 1e-12) / circle_diameter)
            ):
                diameter_sides.append(width)
                thickness_sides.append(height)
            else:
                diameter_sides.append(height)
                thickness_sides.append(width)
    ratios = [thickness / diameter for thickness, diameter in zip(thickness_sides, diameter_sides) if diameter > 1e-9]
    return {
        "n_circles": len(circles),
        "n_rectangles": len(rectangles),
        "circle_diameter": circle_diameter,
        "diameter_sides": diameter_sides,
        "thickness_sides": thickness_sides,
        "thickness_ratios": ratios,
    }


def score_disk_semantics(
    predicted: dict[str, Any], target: dict[str, Any]
) -> dict[str, Any]:
    """Score the one-circle + two D-by-T views of a simple positive disk."""

    pred = _disk_features(predicted)
    gt = _disk_features(target)
    applicable = gt["n_circles"] == 1 and gt["n_rectangles"] == 2 and not target.get("arcs")
    gt_ratios = gt["thickness_ratios"]
    data_valid = bool(
        applicable
        and len(gt_ratios) == 2
        and all(math.isfinite(value) and value > 1e-9 for value in gt_ratios)
    )
    circle_count = _count_f1(pred["n_circles"], gt["n_circles"])
    rectangle_count = _count_f1(pred["n_rectangles"], gt["n_rectangles"])
    completeness = 0.5 * (circle_count + rectangle_count)
    if not applicable:
        return {
            "score": 1.0,
            "applicable": False,
            "data_valid": False,
            "diameter_score": 1.0,
            "thickness_score": 1.0,
            "view_completeness": round(completeness, 6),
        }
    if pred["circle_diameter"] and pred["diameter_sides"]:
        diameter_errors = [
            abs(math.log(max(value, 1e-12) / pred["circle_diameter"]))
            for value in pred["diameter_sides"]
        ]
        diameter_score = float(np.mean(np.exp(-np.asarray(diameter_errors) / DISK_RATIO_DECAY)))
    else:
        diameter_score = 0.0
    if data_valid and pred["thickness_ratios"]:
        pred_ratio = float(np.median(pred["thickness_ratios"]))
        gt_ratio = float(np.median(gt_ratios))
        ratio_quality = math.exp(-abs(math.log(pred_ratio / gt_ratio)) / DISK_RATIO_DECAY)
        internal = float(
            np.mean(
                np.exp(
                    -np.abs(np.log(np.asarray(pred["thickness_ratios"]) / pred_ratio))
                    / DISK_RATIO_DECAY
                )
            )
        )
        thickness_score = ratio_quality * internal
    else:
        pred_ratio = None
        gt_ratio = float(np.median(gt_ratios)) if gt_ratios else None
        thickness_score = 0.0
    score = completeness * (0.25 * diameter_score + 0.75 * thickness_score)
    return {
        "score": round(float(score), 6),
        "applicable": True,
        "data_valid": data_valid,
        "diameter_score": round(float(diameter_score), 6),
        "thickness_score": round(float(thickness_score), 6),
        "view_completeness": round(float(completeness), 6),
        "predicted_thickness_ratio": round(pred_ratio, 8) if pred_ratio is not None else None,
        "target_thickness_ratio": round(gt_ratio, 8) if gt_ratio is not None else None,
        "predicted_features": pred,
        "target_features": gt,
    }


def legacy_dimension_entities(entities: dict[str, Any]) -> dict[str, Any]:
    """Adapt rich v4 dimensions to the v2 tuple API without losing primitives."""

    converted = {
        "lines": list(entities.get("lines", [])),
        "circles": list(entities.get("circles", [])),
        "arcs": list(entities.get("arcs", [])),
        "dims": [],
    }
    for dimension in entities.get("dims", []):
        if not isinstance(dimension, dict):
            converted["dims"].append(dimension)
            continue
        definition = dimension.get("definition_point", (0.0, 0.0))
        binding = dimension.get("binding_point", definition)
        converted["dims"].append(
            (
                int(dimension.get("dimtype", 0)),
                str(dimension.get("text", "")),
                str(dimension.get("layer", "0")),
                float(definition[0]),
                float(definition[1]),
                float(binding[0]),
                float(binding[1]),
            )
        )
    return converted


def representation_quality(raw: dict[str, Any], canonical: dict[str, Any]) -> float:
    """Continuous anti-duplication score that does not punish edge splitting."""

    raw_length = sum(math.dist(line[:2], line[2:4]) for line in raw.get("lines", []))
    union_length = sum(math.dist(line[:2], line[2:4]) for line in canonical.get("lines", []))
    line_quality = min(1.0, union_length / raw_length) if raw_length > 1e-12 else 1.0
    primitive_qualities = [line_quality]
    for family in ("circles", "arcs"):
        values = raw.get(family, [])
        if values:
            keys = {
                tuple(round(float(item), 8) if isinstance(item, (int, float)) else str(item) for item in value)
                for value in values
            }
            primitive_qualities.append(len(keys) / len(values))
    return float(np.mean(primitive_qualities))


__all__ = [
    "canonicalize_entities",
    "extract_entities",
    "legacy_dimension_entities",
    "representation_quality",
    "score_dimension_consistency",
    "score_disk_semantics",
]

	#!/usr/bin/env python3
	"""Representation-invariant metric helpers for the direct ezdxf task.

	Metric v4 is deliberately versioned beside v2/v3. Its first validated scope is
	the DISK family. The important change is that a mathematically straight edge
	is scored as one edge regardless of how many contiguous DXF LINE entities were
	used to encode it. Rich DIMENSION extraction also lets the evaluator verify
	that displayed values agree with measured drawing geometry under one sheet
	scale, instead of checking dimension text alone.
	"""

	from __future__ import annotations

	import math
	from collections import defaultdict
	from pathlib import Path
	from typing import Any, Iterable

	import ezdxf
	import numpy as np

	try: # package import used by tests
	from . import eval_geometric02 as EV2
	from . import eval_geometric03 as EV3
	except ImportError: # direct/module import used by reward workers
	import eval_geometric02 as EV2
	import eval_geometric03 as EV3


	CANONICAL_ABS_TOL = 1e-6
	CANONICAL_REL_TOL = 1e-6
	DIMENSION_SCALE_INLIER_REL = 0.05
	DIMENSION_SCALE_DECAY = 0.05
	DISK_RATIO_DECAY = 0.12


	def _visibility(layer: Any) -> str:
	return "hidden" if "hidden" in str(layer).lower() else "visible"


	def _clean(value: float) -> float:
	value = round(float(value), 12)
	return 0.0 if abs(value) < 5e-13 else value


	def _empty_entities() -> dict[str, list[Any]]:
	return {"lines": [], "circles": [], "arcs": [], "dims": []}


	def extract_entities(dxf_path: str \| Path) -> dict[str, Any] \| None:
	"""Extract primitives plus lossless-enough DIMENSION diagnostics.

	Unlike v2, every dimension is a dictionary containing its raw ezdxf
	measurement and all available definition/reference points. Measurements
	are paper/model-space lengths; displayed text is usually a physical length,
	so their ratio (not equality) is checked by :func:`score_dimension_consistency`.
	"""

	try:
	doc = ezdxf.readfile(str(dxf_path))
	except Exception:
	return None
	entities = _empty_entities()
	for entity in doc.modelspace():
	entity_type = entity.dxftype()
	layer = entity.dxf.layer if entity.dxf.hasattr("layer") else "0"
	if entity_type == "LINE":
	entities["lines"].append(
	(
	float(entity.dxf.start.x),
	float(entity.dxf.start.y),
	float(entity.dxf.end.x),
	float(entity.dxf.end.y),
	layer,
	)
	)
	elif entity_type == "CIRCLE":
	entities["circles"].append(
	(
	float(entity.dxf.center.x),
	float(entity.dxf.center.y),
	float(entity.dxf.radius),
	layer,
	)
	)
	elif entity_type == "ARC":
	entities["arcs"].append(
	(
	float(entity.dxf.center.x),
	float(entity.dxf.center.y),
	float(entity.dxf.radius),
	float(entity.dxf.start_angle),
	float(entity.dxf.end_angle),
	layer,
	)
	)
	elif entity_type == "DIMENSION":
	dimtype = int(entity.dxf.dimtype) if entity.dxf.hasattr("dimtype") else 0
	text = str(entity.dxf.text) if entity.dxf.hasattr("text") else ""
	points: dict[str, tuple[float, float]] = {}
	for name in (
	"defpoint",
	"defpoint2",
	"defpoint3",
	"defpoint4",
	"defpoint5",
	):
	if entity.dxf.hasattr(name):
	point = getattr(entity.dxf, name)
	points[name] = (float(point.x), float(point.y))
	base = dimtype & 0x0F
	if base in (0, 1) and "defpoint2" in points and "defpoint3" in points:
	first, second = points["defpoint2"], points["defpoint3"]
	binding = ((first[0] + second[0]) / 2, (first[1] + second[1]) / 2)
	elif base in (3, 4) and "defpoint4" in points:
	binding = points["defpoint4"]
	else:
	binding = points.get("defpoint", (0.0, 0.0))
	try:
	measurement = float(entity.get_measurement())
	if not math.isfinite(measurement):
	measurement = None
	except Exception:
	measurement = None
	text_family, display_value, _ = EV2.parse_dim_value(text)
	type_family = EV2.dim_type_name(dimtype)
	family = (
	type_family
	if type_family in {"linear", "aligned", "diameter", "radius"}
	else text_family
	)
	entities["dims"].append(
	{
	"dimtype": dimtype,
	"family": family,
	"text": text,
	"display_value": display_value,
	"measurement": measurement,
	"layer": layer,
	"definition_point": points.get("defpoint", (0.0, 0.0)),
	"binding_point": binding,
	"ref_points": points,
	}
	)
	return entities


	def canonicalize_entities(
	entities: dict[str, Any],
	*,
	abs_tol: float = CANONICAL_ABS_TOL,
	rel_tol: float = CANONICAL_REL_TOL,
	) -> dict[str, Any]:
	"""Merge overlapping/touching collinear LINEs without bridging real gaps.

	Visible and hidden geometry are isolated, while arbitrary visible layer
	names are treated as representation detail. The operation is deterministic
	and idempotent. Circles, arcs and rich dimensions are retained unchanged.
	"""

	result = {
	"lines": [],
	"circles": list(entities.get("circles", [])),
	"arcs": list(entities.get("arcs", [])),
	"dims": list(entities.get("dims", [])),
	}
	diagonal = EV2.bbox_diagonal(
	{
	"lines": list(entities.get("lines", [])),
	"circles": result["circles"],
	"arcs": result["arcs"],
	"dims": [],
	}
	)
	tolerance = max(float(abs_tol), float(rel_tol) * max(diagonal, 1.0))
	angular_tolerance = max(1e-10, tolerance / max(diagonal, 1.0))

	clusters: list[dict[str, Any]] = []
	ordered_lines = sorted(
	entities.get("lines", []),
	key=lambda line: (
	_visibility(line[4]),
	min(float(line[0]), float(line[2])),
	min(float(line[1]), float(line[3])),
	max(float(line[0]), float(line[2])),
	max(float(line[1]), float(line[3])),
	str(line[4]),
	),
	)
	for x1, y1, x2, y2, layer in ordered_lines:
	first = np.asarray([float(x1), float(y1)], dtype=float)
	second = np.asarray([float(x2), float(y2)], dtype=float)
	delta = second - first
	length = float(np.linalg.norm(delta))
	if not math.isfinite(length) or length <= 1e-12:
	continue
	direction = delta / length
	if direction[0] < -angular_tolerance or (
	abs(direction[0]) <= angular_tolerance and direction[1] < 0
	):
	direction = -direction
	normal = np.asarray([-direction[1], direction[0]], dtype=float)
	offset = float(np.dot(normal, first))
	visibility = _visibility(layer)
	selected = None
	for cluster in clusters:
	if cluster["visibility"] != visibility:
	continue
	cross = abs(float(np.cross(cluster["direction"], direction)))
	if cross > angular_tolerance:
	continue
	# Recompute offset using the cluster normal so reversed/nearly-equal
	# directions cannot split an otherwise identical support line.
	cluster_offset = float(np.dot(cluster["normal"], first))
	if abs(cluster_offset - cluster["offset"]) <= tolerance:
	selected = cluster
	break
	if selected is None:
	selected = {
	"visibility": visibility,
	"direction": direction,
	"normal": normal,
	"offset": offset,
	"intervals": [],
	"layers": [],
	}
	clusters.append(selected)
	start = float(np.dot(selected["direction"], first))
	end = float(np.dot(selected["direction"], second))
	selected["intervals"].append((min(start, end), max(start, end)))
	selected["layers"].append(str(layer))

	merged: list[tuple[float, float, float, float, str]] = []
	for cluster in clusters:
	intervals = sorted(cluster["intervals"])
	if not intervals:
	continue
	union: list[list[float]] = [[intervals[0][0], intervals[0][1]]]
	for start, end in intervals[1:]:
	if start <= union[-1][1] + tolerance:
	union[-1][1] = max(union[-1][1], end)
	else:
	union.append([start, end])
	output_layer = "HIDDEN" if cluster["visibility"] == "hidden" else min(cluster["layers"])
	for start, end in union:
	first = cluster["direction"] * start + cluster["normal"] * cluster["offset"]
	second = cluster["direction"] * end + cluster["normal"] * cluster["offset"]
	merged.append(
	(
	_clean(first[0]),
	_clean(first[1]),
	_clean(second[0]),
	_clean(second[1]),
	output_layer,
	)
	)
	result["lines"] = sorted(
	merged,
	key=lambda line: (_visibility(line[4]), *line[:4], str(line[4])),
	)
	return result


	def _point_segment_distance(
	point: tuple[float, float], line: tuple[Any, ...]
	) -> float:
	value = np.asarray(point, dtype=float)
	first = np.asarray(line[:2], dtype=float)
	second = np.asarray(line[2:4], dtype=float)
	delta = second - first
	denominator = float(np.dot(delta, delta))
	if denominator <= 1e-18:
	return float(np.linalg.norm(value - first))
	fraction = float(np.dot(value - first, delta) / denominator)
	projection = first + min(1.0, max(0.0, fraction)) * delta
	return float(np.linalg.norm(value - projection))


	def _distance_to_geometry(point: tuple[float, float], entities: dict[str, Any]) -> float:
	distances: list[float] = []
	for line in entities.get("lines", []):
	if _visibility(line[-1]) == "visible":
	distances.append(_point_segment_distance(point, line))
	for cx, cy, radius, layer in entities.get("circles", []):
	if _visibility(layer) == "visible":
	radial = math.dist(point, (float(cx), float(cy)))
	distances.append(abs(radial - abs(float(radius))))
	distances.append(radial) # centre references are valid for circular dims
	for cx, cy, radius, start, end, layer in entities.get("arcs", []):
	if _visibility(layer) == "visible":
	distances.append(abs(math.dist(point, (float(cx), float(cy))) - abs(float(radius))))
	return min(distances) if distances else math.inf


	def _dimension_reference_points(dimension: dict[str, Any]) -> list[tuple[float, float]]:
	refs = dimension.get("ref_points") or {}
	family = str(dimension.get("family", ""))
	if family in {"linear", "aligned"}:
	return [refs[name] for name in ("defpoint2", "defpoint3") if name in refs]
	if family in {"diameter", "radius"}:
	points = [refs[name] for name in ("defpoint4", "defpoint") if name in refs]
	return points or list(refs.values())
	return list(refs.values())


	def score_dimension_consistency(entities: dict[str, Any]) -> dict[str, Any]:
	"""Check shared drawing scale and whether dimensions bind to the outline."""

	dimensions = entities.get("dims", [])
	if not dimensions:
	return {
	"score": 1.0,
	"n_dimensions": 0,
	"n_valid": 0,
	"n_invalid": 0,
	"drawing_scale": None,
	"scale_consistency": 1.0,
	"binding_quality": 1.0,
	"data_valid": True,
	"items": [],
	}
	diagonal = max(EV2.bbox_diagonal(entities), 1.0)
	binding_sigma = max(1e-6, 0.005 * diagonal)
	items: list[dict[str, Any]] = []
	valid_logs: list[float] = []
	for dimension in dimensions:
	value = dimension.get("display_value")
	measurement = dimension.get("measurement")
	family = str(dimension.get("family", ""))
	valid = (
	family in {"linear", "aligned", "diameter", "radius"}
	and isinstance(value, (int, float))
	and isinstance(measurement, (int, float))
	and math.isfinite(float(value))
	and math.isfinite(float(measurement))
	and abs(float(value)) > 1e-9
	and abs(float(measurement)) > 1e-9
	)
	ratio = abs(float(measurement) / float(value)) if valid else None
	if ratio is not None and ratio > 0:
	valid_logs.append(math.log(ratio))
	references = _dimension_reference_points(dimension)
	reference_distances = [_distance_to_geometry(point, entities) for point in references]
	finite_distances = [distance for distance in reference_distances if math.isfinite(distance)]
	if finite_distances:
	# Every extension/reference endpoint should be attached. A mean
	# prevents one valid endpoint from hiding a detached second endpoint.
	binding = float(np.mean(np.exp(-np.asarray(finite_distances) / binding_sigma)))
	else:
	binding = 0.0
	items.append(
	{
	"family": family,
	"text": dimension.get("text", ""),
	"display_value": value,
	"measurement": measurement,
	"scale_ratio": ratio,
	"valid": bool(valid),
	"binding_quality": round(binding, 6),
	"reference_distance_mean": (
	round(float(np.mean(finite_distances)), 6) if finite_distances else None
	),
	}
	)

	if valid_logs:
	threshold = math.log1p(DIMENSION_SCALE_INLIER_REL)
	candidates: list[tuple[int, float, list[float]]] = []
	global_median = float(np.median(valid_logs))
	for candidate in valid_logs:
	inliers = [value for value in valid_logs if abs(value - candidate) <= threshold]
	centre = float(np.median(inliers))
	candidates.append((len(inliers), -abs(centre - global_median), inliers))
	_, _, consensus = max(candidates, key=lambda value: (value[0], value[1]))
	centre_log = float(np.median(consensus))
	qualities = [
	math.exp(-abs(value - centre_log) / DIMENSION_SCALE_DECAY)
	for value in valid_logs
	]
	scale_consistency = float(np.mean(qualities))
	drawing_scale = math.exp(centre_log)
	else:
	centre_log = None
	scale_consistency = 0.0
	drawing_scale = None
	valid_index = 0
	for item in items:
	if item["scale_ratio"] is None or centre_log is None:
	item["scale_inlier"] = False
	item["scale_quality"] = 0.0
	continue
	log_ratio = valid_logs[valid_index]
	valid_index += 1
	item["scale_inlier"] = abs(log_ratio - centre_log) <= math.log1p(DIMENSION_SCALE_INLIER_REL)
	item["scale_quality"] = round(
	math.exp(-abs(log_ratio - centre_log) / DIMENSION_SCALE_DECAY), 6
	)
	binding_quality = float(np.mean([item["binding_quality"] for item in items]))
	n_valid = len(valid_logs)
	n_invalid = len(dimensions) - n_valid
	completeness = n_valid / max(len(dimensions), 1)
	score = scale_consistency * binding_quality * completeness
	return {
	"score": round(float(score), 6),
	"n_dimensions": len(dimensions),
	"n_valid": n_valid,
	"n_invalid": n_invalid,
	"drawing_scale": round(float(drawing_scale), 8) if drawing_scale is not None else None,
	"scale_consistency": round(float(scale_consistency), 6),
	"binding_quality": round(float(binding_quality), 6),
	"data_valid": bool(n_invalid == 0 and scale_consistency >= 0.8),
	"items": items,
	}


	def _connected_line_components(lines: list[tuple[Any, ...]], tolerance: float) -> list[list[tuple[Any, ...]]]:
	if not lines:
	return []
	parent = list(range(len(lines)))

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

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

	endpoints = [
	(np.asarray(line[:2], dtype=float), np.asarray(line[2:4], dtype=float))
	for line in lines
	]
	for first in range(len(lines)):
	for second in range(first + 1, len(lines)):
	if any(
	float(np.linalg.norm(a - b)) <= tolerance
	for a in endpoints[first]
	for b in endpoints[second]
	):
	union(first, second)
	groups: dict[int, list[tuple[Any, ...]]] = defaultdict(list)
	for index, line in enumerate(lines):
	groups[find(index)].append(line)
	return list(groups.values())


	def _rectangle_sides(entities: dict[str, Any]) -> list[tuple[float, float]]:
	visible = [line for line in entities.get("lines", []) if _visibility(line[-1]) == "visible"]
	tolerance = max(1e-6, 1e-5 * max(EV2.bbox_diagonal(entities), 1.0))
	rectangles: list[tuple[float, float]] = []
	for component in _connected_line_components(visible, tolerance):
	points = np.asarray([point for line in component for point in (line[:2], line[2:4])], dtype=float)
	low = points.min(axis=0)
	high = points.max(axis=0)
	width, height = float(high[0] - low[0]), float(high[1] - low[1])
	if width <= tolerance or height <= tolerance:
	continue
	horizontal = vertical = 0.0
	for x1, y1, x2, y2, _ in component:
	if abs(float(y2) - float(y1)) <= tolerance:
	horizontal += abs(float(x2) - float(x1))
	elif abs(float(x2) - float(x1)) <= tolerance:
	vertical += abs(float(y2) - float(y1))
	if horizontal >= 1.8 * width and vertical >= 1.8 * height:
	rectangles.append((width, height))
	return rectangles


	def _count_f1(predicted: int, target: int) -> float:
	if predicted == 0 and target == 0:
	return 1.0
	if predicted == 0 or target == 0:
	return 0.0
	precision = min(predicted, target) / predicted
	recall = min(predicted, target) / target
	return 2 * precision * recall / (precision + recall)


	def _disk_features(entities: dict[str, Any]) -> dict[str, Any]:
	circles = [circle for circle in entities.get("circles", []) if _visibility(circle[-1]) == "visible"]
	rectangles = _rectangle_sides(entities)
	circle_diameter = float(np.median([2 * abs(float(circle[2])) for circle in circles])) if circles else None
	diameter_sides: list[float] = []
	thickness_sides: list[float] = []
	if circle_diameter and circle_diameter > 1e-9:
	for width, height in rectangles:
	if abs(math.log(max(width, 1e-12) / circle_diameter)) <= abs(
	math.log(max(height, 1e-12) / circle_diameter)
	):
	diameter_sides.append(width)
	thickness_sides.append(height)
	else:
	diameter_sides.append(height)
	thickness_sides.append(width)
	ratios = [thickness / diameter for thickness, diameter in zip(thickness_sides, diameter_sides) if diameter > 1e-9]
	return {
	"n_circles": len(circles),
	"n_rectangles": len(rectangles),
	"circle_diameter": circle_diameter,
	"diameter_sides": diameter_sides,
	"thickness_sides": thickness_sides,
	"thickness_ratios": ratios,
	}


	def score_disk_semantics(
	predicted: dict[str, Any], target: dict[str, Any]
	) -> dict[str, Any]:
	"""Score the one-circle + two D-by-T views of a simple positive disk."""

	pred = _disk_features(predicted)
	gt = _disk_features(target)
	applicable = gt["n_circles"] == 1 and gt["n_rectangles"] == 2 and not target.get("arcs")
	gt_ratios = gt["thickness_ratios"]
	data_valid = bool(
	applicable
	and len(gt_ratios) == 2
	and all(math.isfinite(value) and value > 1e-9 for value in gt_ratios)
	)
	circle_count = _count_f1(pred["n_circles"], gt["n_circles"])
	rectangle_count = _count_f1(pred["n_rectangles"], gt["n_rectangles"])
	completeness = 0.5 * (circle_count + rectangle_count)
	if not applicable:
	return {
	"score": 1.0,
	"applicable": False,
	"data_valid": False,
	"diameter_score": 1.0,
	"thickness_score": 1.0,
	"view_completeness": round(completeness, 6),
	}
	if pred["circle_diameter"] and pred["diameter_sides"]:
	diameter_errors = [
	abs(math.log(max(value, 1e-12) / pred["circle_diameter"]))
	for value in pred["diameter_sides"]
	]
	diameter_score = float(np.mean(np.exp(-np.asarray(diameter_errors) / DISK_RATIO_DECAY)))
	else:
	diameter_score = 0.0
	if data_valid and pred["thickness_ratios"]:
	pred_ratio = float(np.median(pred["thickness_ratios"]))
	gt_ratio = float(np.median(gt_ratios))
	ratio_quality = math.exp(-abs(math.log(pred_ratio / gt_ratio)) / DISK_RATIO_DECAY)
	internal = float(
	np.mean(
	np.exp(
	-np.abs(np.log(np.asarray(pred["thickness_ratios"]) / pred_ratio))
	/ DISK_RATIO_DECAY
	)
	)
	)
	thickness_score = ratio_quality * internal
	else:
	pred_ratio = None
	gt_ratio = float(np.median(gt_ratios)) if gt_ratios else None
	thickness_score = 0.0
	score = completeness * (0.25 * diameter_score + 0.75 * thickness_score)
	return {
	"score": round(float(score), 6),
	"applicable": True,
	"data_valid": data_valid,
	"diameter_score": round(float(diameter_score), 6),
	"thickness_score": round(float(thickness_score), 6),
	"view_completeness": round(float(completeness), 6),
	"predicted_thickness_ratio": round(pred_ratio, 8) if pred_ratio is not None else None,
	"target_thickness_ratio": round(gt_ratio, 8) if gt_ratio is not None else None,
	"predicted_features": pred,
	"target_features": gt,
	}


	def legacy_dimension_entities(entities: dict[str, Any]) -> dict[str, Any]:
	"""Adapt rich v4 dimensions to the v2 tuple API without losing primitives."""

	converted = {
	"lines": list(entities.get("lines", [])),
	"circles": list(entities.get("circles", [])),
	"arcs": list(entities.get("arcs", [])),
	"dims": [],
	}
	for dimension in entities.get("dims", []):
	if not isinstance(dimension, dict):
	converted["dims"].append(dimension)
	continue
	definition = dimension.get("definition_point", (0.0, 0.0))
	binding = dimension.get("binding_point", definition)
	converted["dims"].append(
	(
	int(dimension.get("dimtype", 0)),
	str(dimension.get("text", "")),
	str(dimension.get("layer", "0")),
	float(definition[0]),
	float(definition[1]),
	float(binding[0]),
	float(binding[1]),
	)
	)
	return converted


	def representation_quality(raw: dict[str, Any], canonical: dict[str, Any]) -> float:
	"""Continuous anti-duplication score that does not punish edge splitting."""

	raw_length = sum(math.dist(line[:2], line[2:4]) for line in raw.get("lines", []))
	union_length = sum(math.dist(line[:2], line[2:4]) for line in canonical.get("lines", []))
	line_quality = min(1.0, union_length / raw_length) if raw_length > 1e-12 else 1.0
	primitive_qualities = [line_quality]
	for family in ("circles", "arcs"):
	values = raw.get(family, [])
	if values:
	keys = {
	tuple(round(float(item), 8) if isinstance(item, (int, float)) else str(item) for item in value)
	for value in values
	}
	primitive_qualities.append(len(keys) / len(values))
	return float(np.mean(primitive_qualities))


	__all__ = [
	"canonicalize_entities",
	"extract_entities",
	"legacy_dimension_entities",
	"representation_quality",
	"score_dimension_consistency",
	"score_disk_semantics",
	]