每个工业垂直，都需要它自己的 Cosmos。
Every industrial vertical needs its own Cosmos.
不是从视频里学物理，而是把已有的第一性原理模型，包成能批量生成边缘场景、能被 AI Agent 调用的世界模型。调控台（P0–P2）是产品的门面；这里是它为什么会复利——同一套物理，批量产出现实里拿不到的长尾，供下游控制器训练和验证。
Not learning physics from video — wrapping an existing first-principles model into a world model that batch-generates edge cases and is callable by AI agents. The console (P0–P2) is the front door; this is why it compounds — the same physics mass-produces the long tail reality can't give you, to train and validate downstream controllers.
同一套守恒律物理，不只是跑得快——它能批量生成现实里拿不到的长尾场景：故障 × 严重度 × 极端边界 × 一次网的隐性 A↔B 耦合，并自动贴标签（峰值偏差、安全越界、基线 PI 是否失效）。这些就是供下游控制器（PI/MPC/RL）做基准和训练的数据。因为是从守恒律推出来的，它不会幻觉出违反热力学的场景。
The same conservation-law physics doesn't just run fast — it batch-generates the long tail reality can't give you: faults × severities × extreme boundaries × the primary network's hidden A↔B coupling, auto-labeled (peak error, safety violation, whether a baseline PI fails). This is the benchmark/training corpus for downstream controllers (PI/MPC/RL). Derived from conservation laws, it can't hallucinate a thermodynamics-violating scenario.
和调控台是同一个引擎的两端
Two ends of the same engine
左端是操作员的「执行前预演」（plant.ts 从当前态起跑）；右端是这里的「批量造边缘场景」（同一个 plant.ts / network.ts 扫遍故障与边界）。中间用 P2 的决策记录缝合：每次真实执行的残差，既校准模型，又作为带标签的真实尾部数据汇入场景库。
One end is the operator's pre-execution preview (plant.ts from the current state); this end is batch edge-case generation (the same plant.ts / network.ts swept across faults and boundaries). P2's decision records stitch them: every real execution's residual both calibrates the model and joins the library as labeled real-world tail data.