perf: move numeric manager term params to runtime buffers instead of plan_key constants

[wlgys8]

背景

#32 之后 manager kernel 按 plan_key 在进程内去重，但一些纯数值型配置量（termination threshold、obs scale/gain、reward/obs 裸标量参数等）仍以 repr(value) 烘进生成源码与 plan_key，导致调参型变体各自触发一次全额 kernel 重编译。 solves #34。

改动

1. 标量 term 参数 runtime buffer 化（#34 优先级 1）

• 非 kernel-data 的 float 标量参数（observation / reward / termination / sim_reset 四处 dispatch 分支）改为 lowering 进固定 np.float32 的 shared 输入槽（TermScalarBuffer，@kernel_data）：布局指纹进 plan_key，数值运行时读取。
• int/bool 等仍保持编译期常量（多为结构/控制流用途）。
• warmup 值归一化为 np.float32，与 fused kernel 内联特化签名一致。

2. ctrl_dt 移出 plan_key（#34 优先级 2）

• task.reward_weights 存原始权重，dt 缩放移入 evaluate kernel 内 reward * weight * ctx.dt，消除 manager_context_dt 与 reward weights 编译期相乘的双重烘入。

3. 启动日志

• Manager env 每个启动步骤（build plan / compile kernels / term specializations / warmup）单行 INFO 带耗时，并在编译前按三态分类缓存：in-process / disk / miss（miss 时预告首次编译耗时）。
• 内核缓存探测覆盖 numba 两种落盘布局（NUMBA_CACHE_DIR 设置与否），_invalidate_generated_cache 同步修正；顺带修复默认布局下 fallback 清理无法删除缓存文件的问题。
• runner.create_training_handle 各阶段（backend 解析、run context、trainer 就绪）带耗时日志。
• fastsac spawn 的 collector/learner 子进程配置 INFO 日志，启动耗时可见。

性能评测（g1-wbt-dance, 4096 envs）

指标	修改前	修改后
fused evaluate kernel (median)	1.37 ms	1.32 ms（持平）
完整 step (median)	~103 ms	~93 ms（噪声内持平）
合成标量加压微基准（32 个标量参数）	34.8 µs	38.9 µs（+10%，失去常量折叠的最坏情况）
调参变体二次编译	~1.7 s/个	≈ 0（plan 复用）

端到端 step 无可测回退；调参变体的编译开销从线性增长变为 0。

测试

• 新增 test_scalar_term_args_and_ctrl_dt_share_one_compiled_plan：不同 scale/threshold/ctrl_dt 变体共享同一 plan_key 且数值语义各自正确。
• 更新 task.reward_weights 断言（原始权重，不再预乘 dt）。
• motrix_env_core/tests 128 项全部通过；prek run 通过。

Fixes #34

[Copilot]

🟡 Changes recommended

There are a couple of correctness/robustness gaps (a weakened test assertion and subprocess logging setup that may no-op under preconfigured handlers) that should be addressed before approval.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR reduces Numba manager kernel recompilation for tuning-only config variants by lowering float scalar term parameters into fixed np.float32 runtime input buffers (so only layout enters plan_key, not numeric values), and by removing ctrl_dt from the plan key via moving dt scaling into the fused evaluate kernel. It also adds more granular startup/compile timing logs and updates/adds tests to validate plan reuse and correct runtime semantics.

Changes:

• Lower float-like scalar term args into a shared runtime buffer (TermScalarBuffer) to avoid baking numeric values into generated source / plan_key.
• Remove ctrl_dt from plan_key by applying ctx.dt scaling inside the generated reward accumulation code.
• Add startup timing/cache-classification logs across manager env startup and RL runner/fastsac subprocesses; update/add tests for plan reuse and semantics.

File summaries

File	Description
motrix_rl/src/motrix_rl/runner.py	Adds per-stage training startup timing logs around backend resolution, run context creation, and handle readiness.
motrix_rl/src/motrix_rl/fastsac/async_impl/worker.py	Configures spawned worker-process logging so INFO startup logs are visible from collector/learner processes.
motrix_envs/tests/test_wbt_numba.py	Adjusts a plan/read-path reuse test to accommodate additional runtime scalar-buffer sources.
motrix_env_core/tests/test_numba_manager.py	Updates reward weight expectation (no longer pre-multiplied by dt) and adds a plan-sharing test across scalar/ctrl_dt variants.
motrix_env_core/src/motrix_env_core/numba/manager/env.py	Refines manager env startup logging to be step-based with timing and completion markers.
motrix_env_core/src/motrix_env_core/numba/manager/compiler/compiler.py	Implements TermScalarBuffer, shifts float scalar args to runtime buffers, removes dt from plan key, and adds build/compile/specialization timing + cache-classification logging.
motrix_env_core/src/motrix_env_core/numba/manager/compiler/codegen.py	Applies ctx.dt scaling inside the generated weighted reward accumulation.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

It changes core manager-kernel compilation/caching semantics (plan key, runtime inputs, cache invalidation) and should receive final human validation for regressions/perf across environments.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 0 new
• Review effort level: Lite