[PowerX] audit power tests and remove the one tautological file, keeping B200 exporter coverage / 审计功耗测试，删除唯一重言文件并保留 B200 覆盖

[edwingao28]

Description

Audited all 15 test files I added since June: three were already removed in #2799, ten exercise real code with hand-worked expectations and stay, one was tautological. runners/test_kimik3_bh_power.py asserted its own stubs (bundle_server_logs wrote server evidence, the test asserted it back) and compared a dumped yaml fixture to itself. Deleted. Its one real invariant, the B200 dcgm-exporter URI carrying exactly one #, moves into the shared contract test beside GB300 and H200; one stub-echo assertion on the enroot fixture there is also removed.

中文

审计了我自六月以来新增的全部 15 个测试文件：3 个已在 #2799 中删除，10 个跑的是真实代码且期望值为手工推导，予以保留，1 个为重言测试。runners/test_kimik3_bh_power.py 断言的是自己的 stub（bundle_server_logs 写入 server evidence，测试再断言该字符串），并把导出的 yaml fixture 与自身比较。已删除。其中唯一真实的不变量，即 B200 dcgm-exporter URI 只含一个 #，已并入共享契约测试，与 GB300、H200 同列；该文件中一条断言 enroot stub 自身输出的语句同时删除。

Related Issue

None.

Type of Change

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

Checklist

• 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
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /reuse-sweep-run on this PR — n/a

Verification

CI pytest for this path: 476 passed. Breaking the B200 launcher's # separator fails the migrated case.

---

Note

Low Risk
Test-only cleanup and consolidation; no production launcher or runtime behavior changes.

Overview
Removes runners/test_kimik3_bh_power.py, which mostly asserted stub behavior and round-tripped YAML fixtures, and drops it from the test-process-result workflow path filters and pytest command.

The meaningful B200 invariant—dcgm-exporter cold import must use docker://nvcr.io#... with exactly one #—moves into test_exporter_cold_import_uses_nvidia_registry alongside GB300 and H200. That test now parametrizes launch_b200-nscale-slurm.sh, optionally skips cache-path rewriting, pulls in enroot_uri_for_image when present, and stops asserting the stub enroot fixture’s file contents (only the import reference is checked).

^{Reviewed by Cursor Bugbot for commit 589d162. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

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

Additional findings (outside the current diff — GitHub can't attach inline comments there):

• 🟡 runners/test_kimik3_bh_power.py — Deleting this file removes the only coverage for b200/h200 Kimi-K3 power runtime selection (wrong-commit detection, power-pin stamping, power on/off routing to the correct git repo) and for evidence bundling on a failed power run; the replacement in utils/test_gb300_power_official_contract.py only exercises the exporter cold-import '#'-separator invariant, not runtime selection. Fix: add or keep parametrized coverage of the b200/h200 wrong-head/power-routing and failure-evidence-bundling behavior (e.g. alongside the existing gb300-only test_kimi_power_gate_rejects_unsupported_route/test_kimi_power_route_rejects_wrong_commit tests) before removing this file, since gb300 is covered but b200/h200 are not.

  Extended reasoning...

  utils/test_gb300_power_official_contract.py has test_kimi_agentx_route_uses_recipe_power_gate, test_kimi_power_gate_rejects_unsupported_route, and test_kimi_power_route_rejects_wrong_commit, but each is parametrized only on launcher_name='launch_gb300-nv.sh' (lines 398-434). No test anywhere now sources launch_b200-nscale-slurm.sh or launch_h200-dgxc-slurm.sh to check that a wrong git HEAD aborts the run, that the AGENTX_POWER_SRT_SLURM_PIN stamp is written only when power is enabled, or that bundle_server_logs is invoked before a non-zero exit on a failed power run for those two launchers. A future change that breaks wrong-head detection or evidence bundling specifically in launch_b200-nscale-slurm.sh or launch_h200-dgxc-slurm.sh would pass CI silently, whereas before this diff it was caught by test_kimi_power_selects_verified_runtime and test_kimi_failed_power_stages_evidence_before_exit.

  Verification: nit (coverage gap, no production break). The candidate's factual claim holds. The deleted runners/test_kimik3_bh_power.py::test_kimi_power_selects_verified_runtime provided real (not merely tautological) b200/h200 coverage: wrong-head abort ("assert result.returncode != 0; assert not stamp.exists()"), pin stamping gated on power ("stamp.read_text().strip() == 'a'*40" only when power), and…

[edwingao28]

Re the b200/h200 route-coverage suggestion in the review summary: not extending it here. The shared harness slices on USES_AGENTX_POWER / Error: dcgm-power requires; b200 enters via USES_DCGM_POWER with a different error string and h200 nests USES_KIMIK3_POWER with FRAMEWORK=vllm, so covering them means per-launcher bash-slicing branches, the pattern this PR removes. Both routes are verified end-to-end by the reused PR sweeps of #3043 (B200) and #3044 (H200) on main.