修复循环信号与事件采样并接入 LSTP 接触定位，补充八路验证及复用实验

相较上一版 Jacobian 确定性复用更新，本次补齐事件边界一致性、结果两侧采样及接触事件定位；保留已有物性复用和组件力学公式。

- 统一 UD00 信号求值与下一事件查询的绝对时间边界，修复循环边界浮点舍入导致的阶段错位、重复或漏报，并覆盖零时长、多阶段及长周期场景。
- 引入原生输出语义 v2：保留规则网格真实时间，补充内部时间事件和状态事件的左邻及事件后采样，按保存时间、状态和离散模式重放结果。
- 两条代码生成路径均发出 LSTP 接触描述，默认定位间隙过零及非负力模式的力截断；仅在接受事件时更新防重复记录，增加 contactEvents 诊断计数。
- 补充 MASS/LSTP 独立事件实验、八路全曲线与驱动阶段配对评估，以及 Amesim 不连续点输出对照和力差定位报告；MASS 新增释放机制仍保留为独立实验。
- 保存局部 probe、context 访问与回退、shadow replay、R288 real skip/typed replay 及阀门数值尾部诊断工具和报告；未证明净收益的实验不启用为生产默认优化。
- 更新原生运行说明和元件建模规范，补充信号边界、输出语义、接触事件和实验依赖回归测试。

验证：五组专项回归共 34 项全部通过；37 个待提交 Python 文件语法检查通过；git diff --cached --check 通过。

tests/manual/event_phase_comparison.py

@@ -0,0 +1,49 @@
+"""Pair saved samples by forcing phase, never by the size of output errors.
+
+All curves share ONE pair of row indices. No interpolation across jumps, no
+time shifting, and no use of force/pressure agreement to select a sample.
+The reference's nearest saved grid row is authoritative; a different native
+row is allowed only in the tiny output-timestamp roundoff window.
+"""
+from __future__ import annotations
+
+import numpy as np
+
+
+def pair_saved_phases(native_times, reference_times, native_signals, reference_signals, grid, step):
+    nt, rt, grid = map(np.asarray, (native_times, reference_times, grid))
+    ns, rs = map(np.asarray, (native_signals, reference_signals))
+    if ns.shape != (len(nt), rs.shape[1]) or len(rs) != len(rt):
+        raise ValueError('Phase signatures must be rows by the same signal columns.')
+    tolerance = step * 1e-7
+    ni = np.full(len(grid), -1, dtype=int)
+    ri = np.full(len(grid), -1, dtype=int)
+    records = []
+    for j, t in enumerate(grid):
+        candidates = np.arange(np.searchsorted(nt, t-tolerance, side='left'),
+                               np.searchsorted(nt, t+tolerance, side='right'))
+        references = np.arange(np.searchsorted(rt, t-tolerance, side='left'),
+                               np.searchsorted(rt, t+tolerance, side='right'))
+        if not len(candidates) or not len(references):
+            records.append(dict(gridIndex=j, time=float(t), status='missing-saved-sample'))
+            continue
+        # Prefer the later row on an exact tie / duplicate timestamp, matching
+        # the comparison's existing right-side duplicate policy.
+        ref = min(references, key=lambda i: (abs(rt[i]-t), -int(i)))
+        closest = min(candidates, key=lambda i: (abs(nt[i]-t), -int(i)))
+        matches = candidates[np.all(np.isclose(ns[candidates], rs[ref], rtol=1e-12, atol=1e-12), axis=1)]
+        ri[j] = ref
+        if not len(matches):
+            records.append(dict(gridIndex=j, time=float(t), status='unmatched-forcing-phase',
+                                referenceTime=float(rt[ref]), referenceSignals=rs[ref].tolist(),
+                                nativeCandidateTimes=nt[candidates].tolist(),
+                                nativeCandidateSignals=ns[candidates].tolist()))
+            continue
+        chosen = min(matches, key=lambda i: (abs(nt[i]-t), -int(i)))
+        ni[j] = chosen
+        if chosen != closest:
+            records.append(dict(gridIndex=j, time=float(t), status='matched-other-event-side',
+                                platformTime=float(nt[chosen]), referenceTime=float(rt[ref]),
+                                originalPlatformTime=float(nt[closest]),
+                                signals=ns[chosen].tolist(), referenceSignals=rs[ref].tolist()))
+    return ni, ri, records