Files
cdsl-cad/cadfs_to_cdsl/selector_binding.py
T
likang 994d06aaea feat(selector): 增加离线候选遍历与严格回放验证 Demo
- 新增 selector_candidate_demo,移除 provenance intent 后枚举候选 selector
- 对候选分支执行有界重建与严格 STEP 比较
- 仅在候选遍历完整且唯一 strict 通过时生成 selector 映射记录
- 增加 selector 候选搜索、预算限制和记录生成的测试
- 保持生产 selector resolver 不受 Demo 逻辑影响
- 更新 CADFS 能力台账,记录 IMPRINT 派生 profile 的 lineage selector 缺口
2026-09-10 15:12:57 +08:00

359 lines
19 KiB
Python

from __future__ import annotations
import math
from copy import deepcopy
from dataclasses import dataclass
from pathlib import Path
import tempfile
from typing import Any
def _score(expected: dict[str, Any], actual: dict[str, Any]) -> float | None:
scores: list[float] = []
reversed_plane_normal = False
for key in ("center_mm", "circle_center_mm", "start_mm", "end_mm", "normal", "axis_origin_mm", "axis_direction"):
if key in expected:
# plane_offset_mm 与平面方程绑定,必须使用记录平面方程时采用的
# plane_normal。face 的局部采样 normal 在 OCC 中可能与其相反。
use_plane_normal = key == "normal" and "plane_offset_mm" in expected and actual.get("plane_normal") is not None
right = actual.get("plane_normal") if use_plane_normal else actual.get(key)
left = expected[key]
if not isinstance(right, (list, tuple)) or len(left) != len(right): return None
delta = math.sqrt(sum((float(a) - float(b)) ** 2 for a, b in zip(left, right)))
if key in {"normal", "axis_direction"}:
# A source CAP_FACE identifies a geometric plane, not the OCC
# orientation of the resulting face. The two kernels may
# report the same cap with inverse normals, especially for a
# start cap. A rotational-face axis is likewise a geometric
# line, whose direction may be reversed by OCC.
reversed_delta = math.sqrt(sum((float(a) + float(b)) ** 2 for a, b in zip(left, right)))
if key == "normal" and use_plane_normal and reversed_delta < delta: reversed_plane_normal = True
delta = min(delta, reversed_delta)
scores.append(max(0.0, 1.0 - delta / 0.05))
for key in ("radius_mm", "plane_offset_mm"):
if key in expected:
try:
value = float(actual[key])
# 平面偏移是 normal · point。面法向反向时,同一几何平面的
# 有符号偏移也必须同步反号,不能只放宽 normal 的比较。
if key == "plane_offset_mm" and reversed_plane_normal: value = -value
delta = abs(float(expected[key]) - value)
except Exception: return None
scores.append(max(0.0, 1.0 - delta / 0.05))
if "minimum_area_mm2" in expected:
try:
if float(actual["area_mm2"]) + 1e-6 < float(expected["minimum_area_mm2"]): return None
except Exception: return None
if "bbox_mm" in expected:
actual_box = actual.get("bbox_mm")
if not isinstance(actual_box, (list, tuple)) or len(actual_box) != 6: return None
delta = max(abs(float(a) - float(b)) for a, b in zip(expected["bbox_mm"], actual_box))
scores.append(max(0.0, 1.0 - delta / 0.05))
return sum(scores) / len(scores) if scores else 0.0
def bind_selector(kind: str, owner_feature_id: str, geometry: dict[str, Any], records: list[dict[str, Any]], *, minimum_score: float = 0.8) -> dict[str, Any]:
candidates = []
for record in records:
owners = record.get("owner_feature_ids") or [record.get("feature_id")]
if record.get("kind") != kind or owner_feature_id not in owners: continue
score = _score(geometry, record.get("geometry") or {})
if score is not None and score >= minimum_score: candidates.append((score, record))
candidates.sort(key=lambda item: (-item[0], str(item[1].get("record_id"))))
if not candidates: raise ValueError("selector_not_found")
if len(candidates) > 1 and abs(candidates[0][0] - candidates[1][0]) <= 1e-9: raise ValueError("selector_ambiguous")
score, record = candidates[0]
return {"kind": kind, "owner_feature_id": owner_feature_id, "stable_id": record["record_id"], "snapshot_id": record["record_id"], "source": "runtime_snapshot", "confidence": round(score, 6), "geometry": record.get("geometry") or geometry}
def _dot(left: list[float], right: list[float]) -> float: return sum(float(a)*float(b) for a, b in zip(left, right))
def _circle_records(expected: dict[str, Any], records: list[dict[str, Any]]) -> list[dict[str, Any]]:
center = expected.get("source_circle_center_mm"); normal = expected.get("source_plane_normal"); radius = float(expected.get("source_circle_radius_mm") or 0)
if not isinstance(center, list) or not isinstance(normal, list) or radius <= 0: return []
plane_offset = _dot(center, normal); matches = []
for record in records:
geometry = record.get("geometry") or {}
if record.get("kind") != "edge" or geometry.get("curve_type") != "circle": continue
points = [geometry.get("start_mm"), geometry.get("end_mm")]
if not all(isinstance(point, list) and len(point) == 3 for point in points): continue
if any(abs(_dot(point, normal) - plane_offset) > 0.05 for point in points): continue
radial = []
for point in points:
delta = [float(point[i])-float(center[i]) for i in range(3)]; axial = _dot(delta, normal)
radial.append(math.sqrt(max(0.0, sum(value*value for value in delta)-axial*axial)))
if all(abs(value-radius) <= max(0.05, radius*1e-4) for value in radial): matches.append(record)
return matches
def _binding_targets(selector: dict[str, Any]):
components = selector.get("intersection_of")
if isinstance(components, list):
for component in components:
if isinstance(component, dict): yield from _binding_targets(component)
return
yield selector
def _runtime_selector(placeholder: dict[str, Any]) -> dict[str, Any]:
"""Convert a legacy prefix placeholder to the runtime selector contract."""
selector = dict(placeholder)
selector.setdefault("source", "runtime_snapshot")
selector.setdefault("confidence", 1.0)
return selector
def _bound_selector(placeholder: dict[str, Any], records: list[dict[str, Any]]) -> tuple[dict[str, Any], list[dict[str, Any]]]:
bound = []
for record in records:
owners = record.get("owner_feature_ids") or [record.get("feature_id")]
selector = {"kind": placeholder["kind"], "owner_feature_id": str(owners[0]), "stable_id": record["record_id"], "snapshot_id": record["record_id"], "source": "runtime_snapshot", "confidence": 1.0, "geometry": record.get("geometry") or {}}
if placeholder.get("binding_feature_id") is not None: selector["binding_feature_id"] = placeholder["binding_feature_id"]
if placeholder.get("owner_match_required"): selector["owner_match_required"] = True
bound.append(selector)
if placeholder.get("match_mode") == "all":
selector = dict(placeholder); selector["matched_selectors"] = bound
return selector, bound
return bound[0], bound
def _selector_roots(feature: dict[str, Any]) -> list[dict[str, Any]]:
roots = list(feature.get("selectors") or [])
for name in ("end_condition", "reverse_end_condition"):
condition = (feature.get("params") or {}).get(name)
reference = condition.get("reference") if isinstance(condition, dict) else None
if isinstance(reference, dict):
roots.append(reference)
return roots
def _selector_key(selector: dict[str, Any]) -> tuple[Any, ...]:
stable_id = selector.get("stable_id")
if stable_id is not None:
return ("stable_id", str(stable_id))
source = selector.get("output_role_source")
return (
"output_role",
selector.get("owner_feature_id"),
selector.get("kind"),
selector.get("output_role"),
source.get("owner_feature_id") if isinstance(source, dict) else None,
source.get("output_role") if isinstance(source, dict) else None,
)
def _bind_feature_selectors(
feature: dict[str, Any],
*,
registry: Any,
body_id_for_feature: dict[str, str | None],
) -> list[dict[str, Any]]:
"""Bind one feature immediately before it runs in the shared session."""
targets = [target for root in _selector_roots(feature) for target in _binding_targets(root)]
resolved: list[dict[str, Any]] = []
for placeholder in targets:
binding_feature_id = placeholder.get("binding_feature_id")
if binding_feature_id is None:
active_body_id = body_id_for_feature.get("__current__")
else:
if not isinstance(binding_feature_id, str) or binding_feature_id not in body_id_for_feature:
raise ValueError(f"{feature['id']}: selector binding feature is missing or forward")
active_body_id = body_id_for_feature[binding_feature_id]
intent = placeholder.get("selector_intent")
runtime_selector = _runtime_selector(placeholder)
resolution = registry.resolve(runtime_selector, active_body_id=active_body_id)
# Legacy owner-qualified, geometry-free context selectors predate the
# runtime's canonical evidence fields. Their compatibility contract
# permits a unique active context object, but never relaxes a
# provenance, output-role, or instance-locked selector.
if (
resolution.status != "resolved"
and not isinstance(intent, dict)
and not runtime_selector.get("owner_match_required")
and not runtime_selector.get("output_role")
):
fallback = dict(runtime_selector)
fallback.pop("owner_feature_id", None)
resolution = registry.resolve(fallback, active_body_id=active_body_id)
if resolution.status != "resolved":
code = resolution.diagnostic.code if resolution.diagnostic is not None else f"selector_{resolution.status}"
raise ValueError(f"{feature['id']}: {code} during incremental replay")
selected = list(resolution.records or ((resolution.record,) if resolution.record is not None else ()))
if not selected:
raise ValueError(f"{feature['id']}: selector_not_found during incremental replay")
public_records = [record.public_dict() for record in selected]
# Operation-role and provenance selectors stay declarative in bound
# CDSL. A runtime record ID is execution evidence, never their durable
# semantic replacement.
if (isinstance(intent, dict) and intent.get("query_family") != "GEOMETRIC") or placeholder.get("output_role"):
resolved.extend(public_records)
continue
selector, bound_selectors = _bound_selector(placeholder, public_records)
placeholder.clear()
placeholder.update(selector)
resolved.extend(bound_selectors)
selectors = feature.get("selectors") or []
feature["selectors"] = list({_selector_key(selector): selector for selector in selectors}.values())
if feature.get("atomic_id") == "pattern_mirror":
planes = [selector for selector in feature["selectors"] if selector.get("kind") == "plane"]
if len(planes) != 1:
raise ValueError(f"{feature['id']}: mirror plane binding is not unique")
feature.setdefault("params", {})["mirror_plane"] = planes[0]
return resolved
@dataclass
class IncrementalBindingReplay:
"""Bound candidate plus the one session that produced its evidence."""
bound_cdsl: dict[str, Any]
evidence: list[dict[str, Any]]
execution: Any
def bind_and_execute_candidate_selectors(cdsl: dict[str, Any]) -> IncrementalBindingReplay:
"""Bind and execute a CADFS candidate in one ordered kernel replay.
A feature is bound only against topology facts registered by earlier
features in this session. Historical ``binding_feature_id`` values select
a retained snapshot ID, so the binder never needs to re-run a prefix or
recreate an OCC body. Resolver failures retain the live session for the
caller to export its last executable checkpoint.
"""
from engine.cdsl_engine.runtime import prepare_cdsl_execution
bound = deepcopy(cdsl)
execution = prepare_cdsl_execution(bound)
if not execution.analysis.runtime_eligible:
first = next((result for result in execution.analysis.feature_results if not result.executable), None)
code = first.blockers[0].code if first and first.blockers else "runtime_ineligible"
raise ValueError(code)
bound_features = {str(feature.get("id") or ""): feature for feature in bound.get("features") or []}
body_id_for_feature: dict[str, str | None] = {"__current__": None}
evidence: list[dict[str, Any]] = []
for index, node in enumerate(execution.analysis.plan):
feature = node.source_feature
try:
resolved = _bind_feature_selectors(
feature,
registry=execution.session.topology,
body_id_for_feature=body_id_for_feature,
)
public_feature = bound_features[node.feature_id]
public_feature["selectors"] = deepcopy(feature.get("selectors") or [])
public_feature["params"] = deepcopy(feature.get("params") or {})
evidence.append({
"feature_id": node.feature_id,
"prefix_feature_count": index,
"selectors": public_feature["selectors"],
"resolved": resolved,
})
execution.execute_next(strict=True)
body_id_for_feature[node.feature_id] = execution.session.body_id
body_id_for_feature["__current__"] = execution.session.body_id
except Exception as error:
setattr(error, "bound_cdsl", bound)
setattr(error, "selector_binding", evidence)
setattr(error, "incremental_execution", execution)
setattr(error, "failed_feature_id", node.feature_id)
raise
return IncrementalBindingReplay(bound, evidence, execution)
def bind_candidate_selectors(cdsl: dict[str, Any]) -> tuple[dict[str, Any], list[dict[str, Any]]]:
"""Legacy diagnostic binder for incomplete or externally supplied CDSL.
The production CADFS rebuild path uses ``bind_and_execute_candidate_selectors``.
This compatibility entry point deliberately keeps prefix replay explicit
for callers that need to inspect a partial document before it is runtime
eligible. It rehydrates exact exported topology facts and calls the same
resolver; provenance selectors never become geometry guesses here.
"""
from engine.cdsl_engine.runtime import rebuild_cdsl
from engine.cdsl_engine.topology import TopologyRegistry
bound = deepcopy(cdsl)
evidence: list[dict[str, Any]] = []
with tempfile.TemporaryDirectory(prefix="cadfs-bind-") as temporary:
for index, feature in enumerate(bound.get("features") or []):
targets = [target for root in _selector_roots(feature) for target in _binding_targets(root)]
if not targets:
continue
prefix_cache: dict[int, tuple[list[dict[str, Any]], str | None, list[dict[str, Any]]]] = {}
def prefix_snapshot(binding_feature_id: str | None) -> tuple[list[dict[str, Any]], str | None, list[dict[str, Any]]]:
prefix_count = index
if binding_feature_id is not None:
binding_index = next(
(item_index for item_index, item in enumerate(bound["features"][:index]) if item["id"] == binding_feature_id),
None,
)
if binding_index is None:
raise ValueError(f"{feature['id']}: selector binding feature is missing or forward")
prefix_count = binding_index + 1
if prefix_count not in prefix_cache:
prefix = deepcopy(bound)
prefix["features"] = bound["features"][:prefix_count]
if not prefix["features"]:
raise ValueError(f"{feature['id']}: selector has no executable prefix")
report = rebuild_cdsl(
prefix,
Path(temporary) / f"prefix-{index}-{prefix_count}.step",
strict=True,
)
body_id = next(
(item.get("body_id") for item in reversed(report.get("feature_results") or []) if item.get("body_id")),
None,
)
prefix_cache[prefix_count] = (
list(report.get("topology_records") or []),
body_id,
list(report.get("topology_deltas") or []),
)
return prefix_cache[prefix_count]
resolved: list[dict[str, Any]] = []
for placeholder in targets:
records, body_id, topology_deltas = prefix_snapshot(placeholder.get("binding_feature_id"))
registry = TopologyRegistry.from_public_snapshot(records, topology_deltas)
intent = placeholder.get("selector_intent")
runtime_selector = _runtime_selector(placeholder)
resolution = registry.resolve(runtime_selector, active_body_id=body_id)
if (
resolution.status != "resolved"
and not isinstance(intent, dict)
and not runtime_selector.get("owner_match_required")
and not runtime_selector.get("output_role")
):
fallback = dict(runtime_selector)
fallback.pop("owner_feature_id", None)
resolution = registry.resolve(fallback, active_body_id=body_id)
if resolution.status != "resolved":
code = resolution.diagnostic.code if resolution.diagnostic is not None else f"selector_{resolution.status}"
raise ValueError(f"{feature['id']}: {code} after prefix rebuild")
selected = list(resolution.records or ((resolution.record,) if resolution.record is not None else ()))
if not selected:
raise ValueError(f"{feature['id']}: selector_not_found after prefix rebuild")
public_records = [record.public_dict() for record in selected]
if (isinstance(intent, dict) and intent.get("query_family") != "GEOMETRIC") or placeholder.get("output_role"):
resolved.extend(public_records)
continue
selector, bound_selectors = _bound_selector(placeholder, public_records)
placeholder.clear()
placeholder.update(selector)
resolved.extend(bound_selectors)
feature["selectors"] = list({_selector_key(selector): selector for selector in feature.get("selectors") or []}.values())
if feature.get("atomic_id") == "pattern_mirror":
planes = [selector for selector in feature["selectors"] if selector.get("kind") == "plane"]
if len(planes) != 1:
raise ValueError(f"{feature['id']}: mirror plane binding is not unique")
feature.setdefault("params", {})["mirror_plane"] = planes[0]
evidence.append({
"feature_id": feature["id"],
"prefix_feature_count": index,
"selectors": feature["selectors"],
"resolved": resolved,
})
return bound, evidence