[wlgys8]

摘要

Closes #136

本 PR 只修复启动就绪判断和 console 状态表达，不修改 learner 训练/数据搬运调度。

• 启动面板等待所有 learner 实际完成首批更新后再切入训练面板；等待期间显示 initializing first update…。
• 首次更新完成立即发送状态，不等待完整日志间隔；使用本次运行的 training_started 状态，避免把 resume 的历史 update counter 当作本次初始化已完成。
• 有统计窗口但没有新增步数时报告零吞吐量；没有速率测量且不在 replay warmup 时显示 measuring…，不再误显示 warming up。
• 无训练更新的短运行仍可正常退出。

范围

首次 lazy compilation / CUDA Graph 初始化仍然可能耗时。本改动把首批初始化保留在启动阶段明确展示，不消除实际编译或 ring 等待。

根据 review，撤回初版中的 chunk 间 drain、staging stream 同步和相关调度测试。这些 learner 流水线改动需要独立正确性和性能验证，不属于本次状态修复范围。更新预算、strict UTD、ring 消费和 CUDA stream 同步保持原实现。

回归覆盖

• plain / Rich console 区分 replay warmup、尚无速率测量和零吞吐量。
• async 面板等待首轮更新、无更新短运行正常结束、采样停顿报告零吞吐量。

验证

• Console 测试：python -m pytest motrix_rl/tests/test_console.py -q，25 passed。
• Pre-commit 和范围缩小后的验证结果见评论。
• FastSAC pytest 因当前 Torch 无法导入（ImportError: libcudnn.so.9）在收集阶段失败，未执行测试；未安装或修改依赖。
• 真实 GPU / CUDA Graph 启动验收尚未完成。

[Copilot]

🔵 Needs a closer look

The changes alter concurrency-sensitive startup gating in the multi-process async FastSAC trainer, and the PR notes the FastSAC test suite could not be executed and real GPU/CUDA-Graph startup behavior was not verified.

No open issues 🎉

What changed in this PR

This PR addresses issue #136, where the FastSAC async trainer prematurely switched from the startup panel to the training panel before learners finished their first (potentially lazily-compiled) update, and where sampling stalls were misrendered as warming up. The fix is scoped strictly to startup-readiness gating and console status wording; it deliberately does not touch learner training/data-movement scheduling.

Changes:

• Added a per-run training_started flag to learner panel payloads (published immediately on the first completed update, not after a full log interval), and gate the training-panel handoff on all learners reporting it — with a short-run escape so runs that end during replay warmup still shut down cleanly.
• Reworked the panel rate window so a non-empty time span with zero step progress reports a measured 0 env-steps/s instead of None, and disambiguated the None rate as warming up (replay warmup) vs measuring… (no span yet).
• Added regression tests for console throughput wording and for async-panel first-update waiting, zero-throughput-after-stall, and no-update short runs.

File	Description
motrix_rl/​src/​motrix_rl/​console.py	Renders warming up vs measuring… based on the warming flag at all three None-rate sites; comment reworded.
motrix_rl_builtin/​src/​motrix_rl_builtin/​fastsac/​async_impl/​worker.py	Emits training_started in payloads and publishes the first-update payload immediately.
motrix_rl_builtin/​src/​motrix_rl_builtin/​fastsac/​async_impl/​train.py	Gates panel handoff on all learners' first updates, adds a dead-learner break, and reports measured-zero throughput.
motrix_rl/​tests/​test_console.py	Parametrized test covering missing-measurement vs replay-warmup vs zero-throughput wording.
motrix_rl_builtin/​tests/​test_fastsac_live_metrics.py	Adds updates/training_started to the scripted learner and three async-panel regression tests.

_{Review effort: Balanced}