Phase 4A：MJWarp collector CUDA graph candidate hot-path profiling baseline

Summary

#2274 的直接前置已经完成：

• refactor(reset): sink reset composition into UniSim packed transaction #2302 / refactor(reset): migrate or sink ResetStateTransaction host composer #2275：UniSim public packed transaction 成为 reset composition 唯一 owner；
• fix(env): restore G1 reset observation boundaries #2305 / Bug: G1 walk reset owner skips phase observation reset and publishes stale action observation #2303：G1 phase observation reset 与 post-reset last_action ordering 已修复；
• diagnostics(runtime): persist MJWarp CUDA graph state in run manifest #2278 / test(runtime): add CUDA Manager carrier placement contracts #2279：backend graph diagnostics 与 CUDA placement contracts 已建立。

本 issue 只做 Phase 4A profiling / attribution，不实现 collector CUDA graph，不修改 public multi-step contract。

Fixed baseline

• UniLab head：8542371a
• default / ROCm dependency floor：
  • unisim-core>=1.7.17
  • unilab-rl==1.4.12
• profiling 时必须记录：
  • GPU / driver / CUDA runtime；
  • PyTorch / Warp / MJWarp / unisim-core / unilab-rl versions；
  • task / algorithm / num_envs / substeps / action dependency；
  • effective collector step block K；
  • profiler command 与 warmup / measured window；
  • runtime_manifest.backend_diagnostics，特别是 cuda_graph.requested/enabled/disable_reason。

Fixed-K baseline

Current SAC / FlashSAC off-policy owners are semantically K=1:

• training.env_steps_per_sync: 1;
• training.inference_slot_capacity: 1;
• FlashSAC TD algo.algo_params.n_step: 1 is return bootstrapping, not the action-block size.

Therefore the first graph candidate must preserve K=1. Any K>1 public multi-step
carrier requires the Phase 5 action-dependency decision (pre-committed action
sequence versus explicit inference overlap) and cannot be inferred from
updates_per_step, policy_frequency, replay capacity, or MPS mode. K is an
explicit semantic contract, not an implementation tunable.

CUDA MPS attribution matrix

CUDA MPS is already an execution-sharing variable, not merely a deployment
detail. The profiler must collect both modes separately:

1. training.cuda_process_sharing: null — independent learner/collector CUDA contexts;
2. training.cuda_process_sharing: mps — validated single-rank Linux/NVIDIA/MJWarp MPS mode.

Requirements:

• Do not mix null and MPS samples in one latency distribution.
• Use repeated, interleaved arms on the same host and daemon topology where possible.
• Record cuda_process_sharing runtime-manifest evidence alongside backend graph diagnostics.
• Keep learner and collector physical GPU UUID, MPS control pipe/server PID, daemon settings, and active-thread policy if set.
• Explain whether apparent collector gains come from shorter collector work, learner/collector overlap, queueing changes, or reduced context arbitration.
• Treat MPS-on and MPS-off as separate Phase 5 go/no-go evidence; neither result may be extrapolated to the other.

This is required because #2065 / ADR-0013 already showed large E2E changes from
MPS, and CUDA graph replay scheduling/stream ownership can behave differently
under shared execution.

Representative workloads

1. Primary：released-dependency MJWarp / FlashSAC / G1 motion tracking route，对齐 chore(deps): bump unilab-rl to 1.4.12 in the ROCm profile #2301 的 diagnostics baseline。
2. Secondary attribution only：MJWarp / SAC / G1 walk route，用于区分 reset density、G1 walk owner lifecycle 与 motion-tracking lifecycle。

不得把 collector-only microbenchmark 结果当作 E2E acceptance。

Profiling boundary

按 collector lifecycle 采集并归因：

policy action publication
→ action apply / control transform
→ physics step
→ scene tensor read refresh
→ observation
→ reward
→ termination
→ metrics / episode counters
→ events / commands
→ reset mask
→ packed reset
→ post-reset observation
→ replay / inference publication

每个 phase 必须区分：

• wall time；
• GPU kernel time；
• CPU launch / Python dispatch；
• queue wait / sync wait；
• allocation；
• H2D / D2H；
• global synchronize；
• data-dependent Python branch。

Required inventory

在 current main 上重新审计并标注 hot path / interval path / cold path：

• bool(tensor.any())
• .item()
• .cpu() / .tolist() / .numpy()
• masked_select
• nonzero()
• host bool reduction
• temporary allocation
• Python term-by-term manager dispatch
• validation and diagnostics
• event interval / circular buffer branch
• reward / metrics / termination episode publication
• reset count and manual reset branches
• replay / inference publication

输出表格必须包含：

• owner file / function；
• call frequency；
• tensor shape / device；
• sync kind；
• whether inside candidate capture boundary；
• proposed disposition：
  • device-resident counter；
  • fixed-shape pack；
  • interval publication；
  • move outside capture；
  • keep eager；
  • requires Phase 5 public contract。

Deliverables

• Environment / version provenance record.
• Real manifests for null and MPS modes showing backend graph and CUDA process-sharing state.
• Lifecycle phase timing tables for primary and secondary workloads, split by MPS mode.
• Ranked host-sync / allocation / Python-dispatch blockers.
• Candidate capture boundary proposal with explicit inside/outside lists.
• Reconciliation with diagnostics(runtime): persist MJWarp CUDA graph state in run manifest #2278 diagnostics and existing timing attribution.
• Split Phase 4B–4D implementation into small owner-level child issues.
• Phase 5 go / no-go recommendation based on measured collector overhead.

Explicit non-goals

• No CUDA graph implementation.
• No public multi-step API change.
• No removal of semantic logs or episode statistics.
• No backend-private API probing.
• No support / performance claim from stale perf(mjwarp): quantify removing dynamic reset row flow for CUDA graph safety #2285 evidence after Bug: G1 walk reset owner skips phase observation reset and publishes stale action observation #2303.
• No E2E acceptance from collector-only benchmarks.

Definition of done

• Profiling evidence is reproducible from the recorded head, commands, environment, and manifest.
• The effective K is recorded explicitly; the first candidate preserves current K=1 semantics.
• MPS-on and MPS-off attribution are separated and reconciled with runtime-manifest evidence.
• Every known D2H / host sync / dynamic branch in the candidate boundary is either removed from the proposed boundary, assigned to a child issue, or justified as interval/cold path.
• Ranked results identify whether collector外围 overhead is large enough to justify Phase 4 implementation and Phase 5 ADR.
• If expected benefit is not supported by evidence, close this route with a no-go recommendation instead of starting graph MVP.