[PowerX] retain healthy window measurements in audit sidecars / 在审计文件中保留健康窗口测量

[edwingao28]

Description

Retain healthy multinode power measurements in audit sidecars when another window fails. Keep package integrity, aggregate publication, and required-power gates unchanged.

Testing: 343 CPU tests pass, including mixed-window regressions. Retained C2 stays invalid. GPU/hosted CI unrun.

Dependency: None; #3353’s gap policy is excluded.

中文

在其他窗口失败时，于审计 sidecar 中保留健康的多节点功耗测量；包完整性、聚合发布及 required-power 门禁保持不变。

测试： 343 项 CPU 测试通过，包括混合窗口回归。旧 C2 包仍判无效。未运行 GPU 或托管 CI。

依赖： 无；不包含 #3353 的 gap 策略改动。

AI 模型使用说明： 无法核实精确模型/版本。主 agent 和子 agent 负责实现及独立审查。

关联 issue： 无。类型为 bug 修复和文档更新；仅离线审计改动，无需 performance changelog 条目。

AI model disclosure

• Model/version: Exact model/version could not be verified.
• Role: Implementation and independent review, including subagents.

Related Issue

None.

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe)

Checklist

• I have completed the AI model disclosure and kept it current
• I have tested my changes locally
• I have updated documentation if necessary
• For every change that can affect benchmark performance and every recipe addition or modification, I have appended a new entry to the physical end of perf-changelog.yaml and have not edited historical entries (N/A: offline audit only.)
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /use <run_id> (or the legacy /reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Beyond the inline finding, I also checked whether selected_window.power_valid/metrics reporting a healthy window as valid with real energy numbers while the sidecar's top-level power_valid is false is itself a bug — it isn't; that's the documented intent of this change (selected_window fields "do not authorize publication"), so I ruled it out as a separate issue.

Extended reasoning...

The change to infx/results/power/multinode.py splits package-integrity validation from per-window measurement retention in the SRT multinode power audit, touching only offline audit/validation logic (no auth, crypto, or data-exposure surface), with matching doc updates and new parametrized tests. Verified the code directly: audit.per_gpu_energy_j/per_gpu_max_sample_gap_s (lines 1142-1143) are set unconditionally whenever a healthy window is selected, regardless of audit.reasons, contradicting the new doc's claim that "top-level/aggregate energy metrics remain absent" when a sibling window fails — matching the already-reported confirmed inline finding. Separately checked the selected_window.power_valid/metrics exposure (line 1305) against the PR's stated intent and doc wording and confirmed it is working as designed, not an additional bug. Given the confirmed finding is already surfacing as an inline comment, a human should review before merge.