Merge branch 'rescue-layout-opt-detached' into feat/3d_layout_and_visualize

unilabos/layout_optimizer/constraints.py

@@ -16,10 +16,12 @@ from .obb import (
     obb_corners,
     obb_min_distance,
     obb_penetration_depth,
-    segment_intersects_obb,
+    segment_obb_intersection_length,
 )
 
 if TYPE_CHECKING:
+    from typing import Any
+
     from .interfaces import CollisionChecker, ReachabilityChecker
 
 # 归一化默认权重 — 1cm距离违规 ≈ 5°角度违规 的惩罚量级
@@ -296,10 +298,7 @@ def _evaluate_single(
         arm_dev = device_map.get(arm_id)
         target_dev = device_map.get(target_device_id)
 
-        # Distance from target's opening surface center to nearest point on arm OBB.
-        # This naturally enforces orientation: a device facing away has its opening
-        # far from the arm, so it fails reachability without needing a separate
-        # facing penalty.
+        # opening surface center → nearest point on arm OBB
         if arm_dev and target_dev:
             opening_pt = _opening_surface_center(target_dev, target_p)
             arm_corners = obb_corners(
@@ -308,27 +307,30 @@ def _evaluate_single(
             nearest = nearest_point_on_obb(opening_pt[0], opening_pt[1], arm_corners)
             dist = math.sqrt((opening_pt[0] - nearest[0])**2 + (opening_pt[1] - nearest[1])**2)
         else:
+            opening_pt = (target_p.x, target_p.y)
+            nearest = (arm_p.x, arm_p.y)
             dist = _device_distance_center(arm_p, target_p) or 0.0
 
+        # 交叉惩罚始终计算（soft, 不依赖可达性结果）
+        crossing_cost = _crossing_penalty(
+            opening_pt, nearest,
+            arm_id, target_device_id,
+            device_map, placement_map,
+        )
+
         arm_pose = {"x": arm_p.x, "y": arm_p.y, "theta": arm_p.theta}
         target_point = {"x": target_p.x, "y": target_p.y, "z": 0.0}
         target_point["_obb_dist"] = dist
         if not reachability_checker.is_reachable(arm_id, arm_pose, target_point):
             if is_hard and not graduated:
                 return math.inf
-            # Graduated: penalty proportional to overshoot
+            # Graduated: overshoot penalty + crossing cost
             max_reach = reachability_checker.arm_reach.get(arm_id, 2.0)
             overshoot = max(0.0, dist - max_reach)
             w = effective_weight * (HARD_MULTIPLIER if is_hard else 1.0)
-            return w * overshoot * 10.0
+            return w * overshoot * 10.0 + crossing_cost
 
-        # Line-of-sight penalty: penalize if any other device OBB blocks
-        # the path from opening to arm
-        los_cost = _line_of_sight_penalty(
-            arm_id, arm_p, target_device_id, target_p,
-            device_map, placement_map, effective_weight,
-        )
-        return los_cost
+        return crossing_cost
 
     if rule == "prefer_aligned":
         alignment_cost = sum(
@@ -482,23 +484,111 @@ def _opening_surface_center(
     return (world_x, world_y)
 
 
-def _line_of_sight_penalty(
+def evaluate_default_hard_constraints_breakdown(
+    devices: list[Device],
+    placements: list[Placement],
+    lab: Lab,
+    collision_checker: CollisionChecker,
+    *,
+    collision_weight: float = DEFAULT_WEIGHT_DISTANCE * HARD_MULTIPLIER,
+    boundary_weight: float = DEFAULT_WEIGHT_DISTANCE * HARD_MULTIPLIER,
+) -> dict[str, float]:
+    """与 evaluate_default_hard_constraints 逻辑相同，但返回分项明细。"""
+    device_map = {d.id: d for d in devices}
+    collision_cost = 0.0
+    boundary_cost = 0.0
+
+    candidate_pairs = sweep_and_prune_pairs(devices, placements)
+    for i, j in candidate_pairs:
+        di, dj = device_map[placements[i].device_id], device_map[placements[j].device_id]
+        ci = obb_corners(placements[i].x, placements[i].y,
+                         di.bbox[0], di.bbox[1], placements[i].theta)
+        cj = obb_corners(placements[j].x, placements[j].y,
+                         dj.bbox[0], dj.bbox[1], placements[j].theta)
+        depth = obb_penetration_depth(ci, cj)
+        if depth > 0:
+            collision_cost += collision_weight * depth
+
+    for p in placements:
+        dev = device_map[p.device_id]
+        hw, hd = p.rotated_bbox(dev)
+        overshoot = 0.0
+        overshoot += max(0.0, hw - p.x)
+        overshoot += max(0.0, (p.x + hw) - lab.width)
+        overshoot += max(0.0, hd - p.y)
+        overshoot += max(0.0, (p.y + hd) - lab.depth)
+        boundary_cost += boundary_weight * overshoot
+
+    return {
+        "collision": collision_cost,
+        "boundary": boundary_cost,
+        "total": collision_cost + boundary_cost,
+        "collision_weight": collision_weight,
+        "boundary_weight": boundary_weight,
+    }
+
+
+def evaluate_constraints_breakdown(
+    devices: list[Device],
+    placements: list[Placement],
+    lab: Lab,
+    constraints: list[Constraint],
+    collision_checker: CollisionChecker,
+    reachability_checker: ReachabilityChecker | None = None,
+) -> list[dict[str, Any]]:
+    """与 evaluate_constraints 逻辑相同，但返回每条约束的分项明细。"""
+    device_map = {d.id: d for d in devices}
+    placement_map = {p.device_id: p for p in placements}
+
+    results = []
+    for c in constraints:
+        cost = _evaluate_single(
+            c, device_map, placement_map, lab, collision_checker, reachability_checker,
+            graduated=True,
+        )
+        ew = c.weight
+        if c.priority and c.priority in PRIORITY_MULTIPLIERS:
+            ew *= PRIORITY_MULTIPLIERS[c.priority]
+        results.append({
+            "name": _constraint_display_name(c),
+            "rule": c.rule_name,
+            "type": c.type,
+            "cost": cost,
+            "weight": ew,
+        })
+    return results
+
+
+def _constraint_display_name(c: Constraint) -> str:
+    """为约束生成可读的显示名称。"""
+    params = c.params
+    if c.rule_name in (
+        "distance_less_than", "distance_greater_than",
+        "minimize_distance", "maximize_distance",
+    ):
+        return f"{c.rule_name}({params.get('device_a', '?')}, {params.get('device_b', '?')})"
+    if c.rule_name == "reachability":
+        return f"reachability({params.get('arm_id', '?')}, {params.get('target_device_id', '?')})"
+    if c.rule_name == "min_spacing":
+        return f"min_spacing(gap={params.get('min_gap', '?')})"
+    if c.rule_name == "prefer_orientation_mode":
+        return f"prefer_orientation_mode({params.get('mode', '?')})"
+    return c.rule_name
+
+
+def _crossing_penalty(
+    opening_pt: tuple[float, float],
+    arm_nearest_pt: tuple[float, float],
     arm_id: str,
-    arm_p: Placement,
     target_id: str,
-    target_p: Placement,
     device_map: dict[str, Device],
     placement_map: dict[str, Placement],
-    weight: float,
 ) -> float:
-    """Penalty for other devices blocking the line from target to arm center.
+    """交叉惩罚：其他设备 OBB 遮挡 opening→arm 路径的长度加权 penalty。
 
-    For each other device whose OBB intersects the segment (target_center → arm_center),
-    adds a penalty proportional to the weight. This encourages layouts where
-    the arm has a clear path to each target.
+    Soft penalty，权重 = DEFAULT_WEIGHT_DISTANCE * 穿过各遮挡设备 OBB 的线段长度之和。
+    始终生效（不论可达性是否通过），为 DE 提供清晰的梯度信号。
     """
-    p1 = (target_p.x, target_p.y)
-    p2 = (arm_p.x, arm_p.y)
     cost = 0.0
     for dev_id, p in placement_map.items():
         if dev_id == arm_id or dev_id == target_id:
@@ -507,6 +597,6 @@ def _line_of_sight_penalty(
         if dev is None:
             continue
         corners = obb_corners(p.x, p.y, dev.bbox[0], dev.bbox[1], p.theta)
-        if segment_intersects_obb(p1, p2, corners):
-            cost += weight * 2.0  # penalty per blocking device
+        crossing_len = segment_obb_intersection_length(opening_pt, arm_nearest_pt, corners)
+        cost += DEFAULT_WEIGHT_DISTANCE * crossing_len
     return cost

unilabos/layout_optimizer/llm_skill/demo_agent.md

@@ -86,7 +86,8 @@ Build the optimize request using:
 - `constraints`: from Step 3 interpret response
 - `workflow_edges`: from Step 3 interpret response
 - `seeder`: `"compact_outward"` (default)
-- `seeder_overrides`: `{"align_weight": W}` — if user explicitly says "align cardinal", "lock to horizontal/vertical", or "angle locked", use `align_weight: 15.0` instead of default 2.0
+- `seeder_overrides`: generally not needed. Cardinal alignment is handled by the `align_cardinal` intent (generates `prefer_aligned` constraint). Do NOT use `align_weight` in seeder_overrides — it is deprecated.
+- `snap_cardinal`: `false` (default). Set `true` only if user explicitly requests snapping to 0/90/180/270.
 - `run_de`: `true`
 - `maxiter`: `200`
 - `seed`: `42`

unilabos/layout_optimizer/obb.py

@@ -183,6 +183,56 @@ def _point_in_convex(
     return True
 
 
+def segment_obb_intersection_length(
+    p1: tuple[float, float],
+    p2: tuple[float, float],
+    corners: list[tuple[float, float]],
+) -> float:
+    """线段 p1-p2 与 OBB（凸多边形）的交集长度。
+
+    Cyrus-Beck 线段裁剪算法。corners 假定为 CCW 顺序（obb_corners 生成）。
+    无交集返回 0.0。
+    """
+    dx = p2[0] - p1[0]
+    dy = p2[1] - p1[1]
+    seg_len_sq = dx * dx + dy * dy
+    if seg_len_sq < 1e-24:
+        return 0.0
+
+    t_enter = 0.0
+    t_exit = 1.0
+    n = len(corners)
+
+    for i in range(n):
+        ax, ay = corners[i]
+        bx, by = corners[(i + 1) % n]
+        # CCW 多边形边的外法线: (ey, -ex), e = b - a
+        ex, ey = bx - ax, by - ay
+        nx, ny = ey, -ex
+
+        denom = nx * dx + ny * dy
+        numer = nx * (p1[0] - ax) + ny * (p1[1] - ay)
+
+        if abs(denom) < 1e-12:
+            if numer > 0:
+                return 0.0  # 在此边外侧且平行
+            continue
+
+        t = -numer / denom
+        if denom < 0:
+            t_enter = max(t_enter, t)   # 进入
+        else:
+            t_exit = min(t_exit, t)     # 退出
+
+        if t_enter > t_exit:
+            return 0.0
+
+    if t_enter >= t_exit:
+        return 0.0
+
+    return (t_exit - t_enter) * math.sqrt(seg_len_sq)
+
+
 def obb_min_distance(
     corners_a: list[tuple[float, float]],
     corners_b: list[tuple[float, float]],

unilabos/layout_optimizer/optimizer.py

@@ -13,7 +13,12 @@ from typing import Any, Callable
 
 import numpy as np
 
-from .constraints import evaluate_constraints, evaluate_default_hard_constraints
+from .constraints import (
+    evaluate_constraints,
+    evaluate_constraints_breakdown,
+    evaluate_default_hard_constraints,
+    evaluate_default_hard_constraints_breakdown,
+)
 from .mock_checkers import MockCollisionChecker, MockReachabilityChecker
 from .models import Constraint, Device, Lab, Placement
 from .pencil_integration import generate_initial_layout
@@ -34,6 +39,7 @@ def _run_de(
     seed: int | None,
     n_devices: int,
     strategy: str = "currenttobest1bin",
+    progress_callback: Callable[[int, np.ndarray, float], None] | None = None,
 ) -> tuple[np.ndarray, float, int]:
     """自定义差分进化循环。
 
@@ -56,6 +62,7 @@ def _run_de(
         seed: 随机种子
         n_devices: 设备数量（用于 per-device crossover）
         strategy: 变异策略，"currenttobest1bin" 或 "best1bin"
+        progress_callback: 每 10 代调用一次 (gen, best_vector, best_cost)
 
     Returns:
         (best_vector, best_cost, n_generations)
@@ -125,6 +132,10 @@ def _run_de(
         # 更新 best_idx（种群可能整体更新）
         best_idx = int(np.argmin(costs))
 
+        # 进度回调：每 10 代报告最优个体状态
+        if progress_callback and gen % 10 == 0:
+            progress_callback(gen, best_vector, best_cost)
+
         # Early stopping：最近 patience 代改善 < 0.1%
         best_cost_history.append(best_cost)
         if len(best_cost_history) >= patience:
@@ -296,6 +307,36 @@ def optimize(
         n, 3 * n, pop_count, maxiter, strategy,
     )
 
+    # DEBUG 模式进度回调：每 10 代输出完整约束分项表格
+    def _progress_cb(gen: int, best_vec: np.ndarray, best_cost_val: float) -> None:
+        if not logger.isEnabledFor(logging.DEBUG):
+            return
+        pls = _vector_to_placements(best_vec, devices)
+        hard_bd = evaluate_default_hard_constraints_breakdown(
+            devices, pls, lab, collision_checker,
+        )
+        lines = [f"=== DE Gen {gen} | best_cost={best_cost_val:.4f} ==="]
+        lines.append(f"  {'Constraint':<45} {'Type':<6} {'Weight':>8} {'Cost':>10}")
+        lines.append(f"  {'─' * 71}")
+        lines.append(
+            f"  {'[predefined] collision':<45} {'hard':<6} {hard_bd['collision_weight']:>8.0f} {hard_bd['collision']:>10.4f}"
+        )
+        lines.append(
+            f"  {'[predefined] boundary':<45} {'hard':<6} {hard_bd['boundary_weight']:>8.0f} {hard_bd['boundary']:>10.4f}"
+        )
+        if constraints:
+            user_bd = evaluate_constraints_breakdown(
+                devices, pls, lab, constraints,
+                collision_checker, reachability_checker,
+            )
+            for item in user_bd:
+                lines.append(
+                    f"  {item['name']:<45} {item['type']:<6} {item['weight']:>8.1f} {item['cost']:>10.4f}"
+                )
+        lines.append(f"  {'─' * 71}")
+        lines.append(f"  {'TOTAL':<45} {'':6} {'':>8} {best_cost_val:>10.4f}")
+        logger.debug("\n".join(lines))
+
     best_vector, best_cost, n_generations = _run_de(
         cost_fn=cost_function,
         bounds=bounds_array,
@@ -308,15 +349,52 @@ def optimize(
         seed=seed,
         n_devices=n,
         strategy=strategy,
+        progress_callback=_progress_cb,
     )
 
     # 评估次数估算：每代 pop_count 次（初始 + 每代 trial）
     n_evaluations = pop_count + n_generations * pop_count
 
-    logger.info(
-        "DE optimization complete: success=%s, cost=%.4f, iterations=%d, evaluations=%d",
-        best_cost < 1e17, best_cost, n_generations, n_evaluations,
+    # 最终布局分项明细（INFO 级别）
+    final_placements = _vector_to_placements(best_vector, devices)
+    hard_bd = evaluate_default_hard_constraints_breakdown(
+        devices, final_placements, lab, collision_checker,
     )
+    # success = 所有 hard 约束均满足（predefined + 用户 hard）
+    all_hard_met = hard_bd["total"] == 0.0
+    # 所有约束的 top violators 候选池（predefined + user）
+    all_violators: list[dict] = [
+        {"name": "[predefined] collision", "cost": hard_bd["collision"]},
+        {"name": "[predefined] boundary", "cost": hard_bd["boundary"]},
+    ]
+    if constraints:
+        user_bd = evaluate_constraints_breakdown(
+            devices, final_placements, lab, constraints,
+            collision_checker, reachability_checker,
+        )
+        user_total = sum(item["cost"] for item in user_bd)
+        for c_item in user_bd:
+            all_violators.append({"name": c_item["name"], "cost": c_item["cost"]})
+            if c_item["type"] == "hard" and c_item["cost"] > 0:
+                all_hard_met = False
+    else:
+        user_bd = []
+        user_total = 0.0
+
+    summary = [
+        "DE complete: success=%s, cost=%.4f, %d gens, %d evals"
+        % (all_hard_met, best_cost, n_generations, n_evaluations),
+        "  Predefined: subtotal=%.4f" % hard_bd["total"],
+    ]
+    if constraints:
+        summary.append(f"  User: subtotal={user_total:.4f}")
+    top_violators = sorted(all_violators, key=lambda x: x["cost"], reverse=True)[:3]
+    top_violators = [v for v in top_violators if v["cost"] > 0]
+    if top_violators:
+        summary.append("  Top violators:")
+        for v in top_violators:
+            summary.append(f"    {v['name']} = {v['cost']:.4f}")
+    logger.info("\n".join(summary))
 
     return _vector_to_placements(best_vector, devices)
 

unilabos/layout_optimizer/server.py

@@ -4,7 +4,14 @@
 集成阶段合并到 Uni-Lab-OS 的 FastAPI 服务中。
 
 运行方式：
-    uvicorn layout_optimizer.server:app --host 0.0.0.0 --port 8000 --reload
+    uvicorn unilabos.layout_optimizer.server:app --host 0.0.0.0 --port 8000 --reload
+
+调试模式（启用 DEBUG 日志，含优化器逐代 cost 明细）：
+    LAYOUT_DEBUG=1 uvicorn unilabos.layout_optimizer.server:app --host 0.0.0.0 --port 8000 --reload
+
+日志文件：
+    自动写入 layout_optimizer/logs/{YYYYMMDD_HHMMSS}.log（始终 DEBUG 级别）。
+    前端 1s 轮询的 GET /scene/placements 200 行不写入日志文件。
 
 前端访问：
     http://localhost:8000/
@@ -13,8 +20,10 @@
 from __future__ import annotations
 
 import logging
+import logging.handlers
 import math
 import os
+from datetime import datetime
 from pathlib import Path
 
 from fastapi import FastAPI
@@ -36,9 +45,40 @@ from .intent_interpreter import InterpretResult, interpret_intents
 from .models import Constraint, Intent
 from .optimizer import optimize
 
-logging.basicConfig(level=logging.INFO)
+_console_level = logging.DEBUG if os.getenv("LAYOUT_DEBUG") else logging.INFO
+# root logger must be DEBUG so the file handler receives all records;
+# console output level is controlled separately via its handler.
+logging.basicConfig(level=logging.DEBUG)
+# basicConfig creates a default StreamHandler — set its level to the console level
+for _h in logging.getLogger().handlers:
+    if isinstance(_h, logging.StreamHandler):
+        _h.setLevel(_console_level)
 logger = logging.getLogger(__name__)
 
+# --- 文件日志：实时写入 logs/ 目录，按启动时间命名 ---
+_LOG_DIR = Path(__file__).parent / "logs"
+_LOG_DIR.mkdir(exist_ok=True)
+_log_file = _LOG_DIR / f"{datetime.now():%Y%m%d_%H%M%S}.log"
+
+
+class _PollingFilter(logging.Filter):
+    """过滤掉前端 1s 轮询产生的 GET /scene/placements 日志行。"""
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        msg = record.getMessage()
+        if "GET /scene/placements" in msg and "200" in msg:
+            return False
+        return True
+
+
+_file_handler = logging.FileHandler(_log_file, encoding="utf-8")
+_file_handler.setLevel(logging.DEBUG)
+_file_handler.setFormatter(
+    logging.Formatter("%(asctime)s %(levelname)-5s [%(name)s] %(message)s")
+)
+_file_handler.addFilter(_PollingFilter())
+logging.getLogger().addHandler(_file_handler)
+
 STATIC_DIR = Path(__file__).parent / "static"
 
 # 可配置路径
@@ -368,6 +408,7 @@ class OptimizeRequest(BaseModel):
     workflow_edges: list[list[str]] = []
     maxiter: int = 200
     seed: int | None = None
+    snap_cardinal: bool = False
 
 
 class PositionXYZ(BaseModel):
@@ -459,8 +500,8 @@ async def run_optimize(request: OptimizeRequest):
                 params={"mode": orientation_mode},
                 weight=request.seeder_overrides.get("orientation_weight", DEFAULT_WEIGHT_ANGLE),
             ))
-        # prefer_aligned: penalize non-cardinal angles
-        align_weight = request.seeder_overrides.get("align_weight", DEFAULT_WEIGHT_ANGLE)
+        # prefer_aligned: penalize non-cardinal angles（默认关闭，用户可通过 align_cardinal intent 或 seeder_overrides 开启）
+        align_weight = request.seeder_overrides.get("align_weight", 0)
         if align_weight > 0:
             constraints.append(Constraint(
                 type="soft",
@@ -486,8 +527,9 @@ async def run_optimize(request: OptimizeRequest):
     else:
         result_placements = seed_placements
 
-    # 5. θ snap post-processing（碰撞安全：snap 后验证，失败则回退）
-    result_placements = snap_theta_safe(result_placements, devices, lab, checker)
+    # 5. θ snap post-processing（opt-in，默认关闭）
+    if request.snap_cardinal:
+        result_placements = snap_theta_safe(result_placements, devices, lab, checker)
 
     # 6. Evaluate final cost (binary mode for pass/fail reporting)
     final_cost = evaluate_default_hard_constraints(

unilabos/layout_optimizer/tests/test_bugfixes_v2.py

@@ -376,3 +376,58 @@ class TestScenarios:
         assert not _has_collision(devices, result)
         for i, p in enumerate(result):
             assert _facing_dot(p, devices[i], lab) > 0
+
+
+# ── V2 Stage 1: 默认关闭 cardinal snap/alignment ────────
+
+class TestV2Stage1Bugfixes:
+    """align_weight 默认为 0，snap_cardinal 默认关闭。"""
+
+    def test_default_align_weight_is_zero(self):
+        """Default request (no seeder_overrides) should NOT inject prefer_aligned."""
+        from fastapi.testclient import TestClient
+        from ..server import app
+
+        client = TestClient(app)
+        resp = client.post("/optimize", json={
+            "devices": [{"id": "opentrons_liquid_handler", "uuid": "u1"}],
+            "lab": {"width": 3, "depth": 3},
+            "seeder": "compact_outward",
+            "run_de": True,
+            "maxiter": 50,
+            "seed": 42,
+        })
+        assert resp.status_code == 200
+
+    def test_snap_cardinal_off_by_default(self):
+        """Default request should NOT snap theta to cardinal."""
+        from fastapi.testclient import TestClient
+        from ..server import app
+
+        client = TestClient(app)
+        resp = client.post("/optimize", json={
+            "devices": [{"id": "opentrons_liquid_handler", "uuid": "u1"}],
+            "lab": {"width": 3, "depth": 3},
+            "seeder": "compact_outward",
+            "run_de": True,
+            "maxiter": 10,
+            "seed": 42,
+        })
+        assert resp.status_code == 200
+
+    def test_snap_cardinal_opt_in(self):
+        """snap_cardinal=True should be accepted and snap angles."""
+        from fastapi.testclient import TestClient
+        from ..server import app
+
+        client = TestClient(app)
+        resp = client.post("/optimize", json={
+            "devices": [{"id": "opentrons_liquid_handler", "uuid": "u1"}],
+            "lab": {"width": 3, "depth": 3},
+            "seeder": "compact_outward",
+            "snap_cardinal": True,
+            "run_de": True,
+            "maxiter": 10,
+            "seed": 42,
+        })
+        assert resp.status_code == 200

unilabos/layout_optimizer/tests/test_constraints.py

@@ -5,11 +5,15 @@ import math
 import pytest
 
 from ..constraints import (
+    _crossing_penalty,
+    _opening_surface_center,
+    DEFAULT_WEIGHT_DISTANCE,
     evaluate_constraints,
     evaluate_default_hard_constraints,
 )
 from ..mock_checkers import MockCollisionChecker, MockReachabilityChecker
-from ..models import Constraint, Device, Lab, Placement
+from ..models import Constraint, Device, Opening, Placement, Lab
+from ..obb import nearest_point_on_obb, obb_corners
 
 
 def _make_devices():
@@ -421,3 +425,81 @@ class TestGraduatedHardConstraints:
             devices, placements, _make_lab(), constraints, checker,
         )
         assert not math.isinf(cost)
+
+
+class TestCrossingPenalty:
+    """_crossing_penalty: 交叉长度加权的 soft penalty。"""
+
+    def _make_device(self, dev_id, bbox=(0.5, 0.5), direction=(0.0, -1.0)):
+        return Device(
+            id=dev_id, name=dev_id, device_type="static",
+            bbox=bbox, height=0.3,
+            openings=[Opening(direction=direction, label="front")],
+        )
+
+    def test_no_blockers_returns_zero(self):
+        """arm 与 target 之间无遮挡设备 → 交叉代价为 0。"""
+        arm = self._make_device("arm", bbox=(2.14, 0.35))
+        target = self._make_device("target")
+        arm_p = Placement(device_id="arm", x=2.0, y=1.0, theta=0.0)
+        target_p = Placement(device_id="target", x=0.5, y=1.0, theta=3.14159)
+        device_map = {"arm": arm, "target": target}
+        placement_map = {"arm": arm_p, "target": target_p}
+
+        opening_pt = _opening_surface_center(target, target_p)
+        arm_corners = obb_corners(arm_p.x, arm_p.y, arm.bbox[0], arm.bbox[1], arm_p.theta)
+        nearest = nearest_point_on_obb(opening_pt[0], opening_pt[1], arm_corners)
+
+        cost = _crossing_penalty(
+            opening_pt, nearest,
+            "arm", "target",
+            device_map, placement_map,
+        )
+        assert cost == 0.0
+
+    def test_one_blocker_proportional_to_length(self):
+        """一个遮挡设备 → cost = DEFAULT_WEIGHT_DISTANCE * 穿过长度。"""
+        arm = self._make_device("arm", bbox=(2.14, 0.35))
+        target = self._make_device("target")
+        blocker = self._make_device("blocker", bbox=(0.5, 0.5))
+        arm_p = Placement(device_id="arm", x=3.0, y=1.0, theta=0.0)
+        target_p = Placement(device_id="target", x=0.0, y=1.0, theta=0.0)
+        blocker_p = Placement(device_id="blocker", x=1.5, y=1.0, theta=0.0)
+        device_map = {"arm": arm, "target": target, "blocker": blocker}
+        placement_map = {"arm": arm_p, "target": target_p, "blocker": blocker_p}
+
+        opening_pt = _opening_surface_center(target, target_p)
+        arm_corners = obb_corners(arm_p.x, arm_p.y, arm.bbox[0], arm.bbox[1], arm_p.theta)
+        nearest = nearest_point_on_obb(opening_pt[0], opening_pt[1], arm_corners)
+
+        cost = _crossing_penalty(
+            opening_pt, nearest,
+            "arm", "target",
+            device_map, placement_map,
+        )
+        # blocker 宽 0.5m，theta=0，路径水平 → 穿过长度 ≈ 0.5m
+        # cost = DEFAULT_WEIGHT_DISTANCE * 0.5 = 100 * 0.5 = 50
+        assert cost > 0
+        assert abs(cost - DEFAULT_WEIGHT_DISTANCE * 0.5) < DEFAULT_WEIGHT_DISTANCE * 0.1
+
+    def test_blocker_off_path_returns_zero(self):
+        """不在路径上的设备 → 交叉代价为 0。"""
+        arm = self._make_device("arm", bbox=(2.14, 0.35))
+        target = self._make_device("target")
+        bystander = self._make_device("bystander", bbox=(0.5, 0.5))
+        arm_p = Placement(device_id="arm", x=3.0, y=1.0, theta=0.0)
+        target_p = Placement(device_id="target", x=0.0, y=1.0, theta=0.0)
+        bystander_p = Placement(device_id="bystander", x=1.5, y=3.0, theta=0.0)
+        device_map = {"arm": arm, "target": target, "bystander": bystander}
+        placement_map = {"arm": arm_p, "target": target_p, "bystander": bystander_p}
+
+        opening_pt = _opening_surface_center(target, target_p)
+        arm_corners = obb_corners(arm_p.x, arm_p.y, arm.bbox[0], arm.bbox[1], arm_p.theta)
+        nearest = nearest_point_on_obb(opening_pt[0], opening_pt[1], arm_corners)
+
+        cost = _crossing_penalty(
+            opening_pt, nearest,
+            "arm", "target",
+            device_map, placement_map,
+        )
+        assert cost == 0.0

unilabos/layout_optimizer/tests/test_e2e_pcr_pipeline.py

@@ -199,6 +199,8 @@ class TestStage3VerifyPlacements:
             "run_de": True,
             "maxiter": 100,
             "seed": 42,
+            "snap_cardinal": True,
+            "seeder_overrides": {"align_weight": 60},
         })
         data = optimize_resp.json()
         assert data["success"] is True

unilabos/layout_optimizer/tests/test_obb.py

@@ -1,7 +1,7 @@
 """Tests for OBB (Oriented Bounding Box) geometry utilities."""
 import math
 import pytest
-from ..obb import obb_corners, obb_overlap, obb_min_distance
+from ..obb import obb_corners, obb_overlap, obb_min_distance, segment_obb_intersection_length
 
 
 class TestObbCorners:
@@ -115,3 +115,42 @@ class TestObbMinDistance:
         a = obb_corners(0, 0, 2.0, 2.0, 0.0)
         b = obb_corners(2.0, 0, 2.0, 2.0, 0.0)
         assert obb_min_distance(a, b) == pytest.approx(0.0)
+
+
+class TestSegmentOBBIntersectionLength:
+    """segment_obb_intersection_length: Cyrus-Beck clipping."""
+
+    def test_segment_fully_outside(self):
+        corners = obb_corners(0, 0, 2, 2, 0)
+        length = segment_obb_intersection_length((-5, 3), (5, 3), corners)
+        assert length == 0.0
+
+    def test_segment_fully_inside(self):
+        corners = obb_corners(0, 0, 4, 4, 0)
+        length = segment_obb_intersection_length((-0.5, 0), (0.5, 0), corners)
+        assert abs(length - 1.0) < 1e-6
+
+    def test_segment_crosses_through(self):
+        corners = obb_corners(0, 0, 2, 2, 0)
+        length = segment_obb_intersection_length((-5, 0), (5, 0), corners)
+        assert abs(length - 2.0) < 1e-6
+
+    def test_segment_partial_overlap(self):
+        corners = obb_corners(0, 0, 2, 2, 0)
+        length = segment_obb_intersection_length((0, 0), (5, 0), corners)
+        assert abs(length - 1.0) < 1e-6
+
+    def test_rotated_obb(self):
+        corners = obb_corners(0, 0, 2, 2, math.pi / 4)
+        length = segment_obb_intersection_length((-3, 0), (3, 0), corners)
+        expected = 2 * math.sqrt(2)
+        assert abs(length - expected) < 1e-4
+
+    def test_zero_length_segment(self):
+        corners = obb_corners(0, 0, 2, 2, 0)
+        assert segment_obb_intersection_length((0, 0), (0, 0), corners) == 0.0
+
+    def test_parallel_outside(self):
+        corners = obb_corners(0, 0, 2, 2, 0)
+        length = segment_obb_intersection_length((-5, 2), (5, 2), corners)
+        assert length == 0.0