refactor: share AgentX matrix generation / 共用 AgentX 矩阵生成逻辑

docs/architecture.md

@@ -244,6 +244,8 @@ Use the master entry to answer what should be benchmarked. Use runner config for
 
 Derived concurrency points, eval selection, topology defaults, names, and runner-derived facts belong in `generate_sweep_configs.py`. Workflows should forward matrix fields, not reimplement generator policy in expressions or shell.
 
+The `full-sweep` and `test-config` commands share fixed-sequence and AgentX row builders. Command-specific selection remains in the callers; the AgentX builder owns worker defaults, offload budgets, experiment names, node counts, and validation. It validates topology and offload budgets before filtering concurrency, preserves point and runner ordering, and filters AgentX bounds without inventing a capped concurrency point.
+
 ### Trigger selection is separate from configuration
 
 `perf-changelog.yaml` selects work and records why. It does not redefine a master entry. This makes the configuration catalog reusable while keeping a reviewable history of what each sweep intended to run.

docs/architecture_zh.md

@@ -244,6 +244,8 @@ bash ./runners/launch_${RUNNER_NAME%%_*}.sh
 
 派生并发点、评测选择、拓扑默认值、名称和运行器派生信息属于 `generate_sweep_configs.py`。工作流应转发矩阵字段，而不应在表达式或 Shell 中重新实现生成器策略。
 
+`full-sweep` 和 `test-config` 命令共用固定序列与 AgentX 的矩阵行构建逻辑。各命令的选择规则仍由调用方负责；AgentX 构建逻辑负责 worker 默认值、卸载预算、实验名称、节点数和验证。它在过滤并发度前验证拓扑与卸载预算，保留并发点和运行器的原有顺序，并仅按上下限过滤 AgentX 并发点，不会额外生成截断到上限的并发点。
+
 ### 触发选择与配置彼此独立
 
 `perf-changelog.yaml` 选择工作并记录原因。它不会重新定义主条目。这样既可以复用配置目录，又能保留可供审查的历史记录，说明每次扫描打算运行什么。