fix: recognize git -c in SRT clone test stub / fix: 让 SRT clone 测试夹具识别 git -c

[edwingao28]

Summary

Clears the InferenceX #3379 CPU Tests failure with the smallest justified fix already landed on main in #3443.

The CI failure was not caused by the B300 native SRT power coverage diff. It is launcher/test contract drift from #3410 (git -c advice.detachedHead=false clone ...) that the PR head absorbed when it merged then-main (b7bff8bc) before #3443 updated the git stub. The stub only matched $1 == clone, so staging never created checkout/, and every test_pool_launcher_stages_artifacts_and_propagates_failure[...] case failed (missing checkout / point-identity.json).

This draft repair PR targets owned branch fix/qwen35-b300-powerx-coverage so ownership of #3379 is preserved.

Root cause

Item	Value
Failing CI	Tests job 108227688377
Checked-out merge	PR head 5c9fc7e6 into then-base b7bff8bc (merge SHA b61f0c91)
Failure remains on current head?	Yes — reproduced locally at 5c9fc7e6
Cause of PR?	No — main-side drift (#3410) surfaced on the merge; fix already on current main (#3443)
Qwen 144/48 acceptance	Unaffected; this is a CPU test-fixture regression, not a measurement defect

Fix

One-line change in utils/test_srt_single_node.py git stub:

# before
if [[ "$1" == clone ]]; then ...
# after (same as main #3443)
if [[ " $* " == *" clone "* ]]; then ...

Verification (CPU-only)

# BEFORE (at 5c9fc7e6)
uv run --locked python -m pytest \
  'utils/test_srt_single_node.py::test_pool_launcher_stages_artifacts_and_propagates_failure[b300-dsxe-none]' \
  'utils/test_srt_single_node.py::test_pool_launcher_stages_artifacts_and_propagates_failure[h200-dgxc-slurm-none]' \
  -v --tb=line
# -> 2 failed (cd checkout: No such file or directory)

# AFTER (this branch)
uv run --locked python -m pytest \
  'utils/test_srt_single_node.py::test_pool_launcher_stages_artifacts_and_propagates_failure' \
  -v --tb=line
# -> 26 passed

No GPU runs. No remote CI reruns for diagnosis.

Main conflict status

Owned PR #3379 remains DIRTY/CONFLICTING vs current main (large project layout migration #3525+). This repair does not resolve that rebase; it only clears the CPU Tests contract failure on the current #3379 lineage. After merging this into #3379, a separate main reconcile is still required before #3379 can be a merge candidate.

AI model disclosure

• gpt-5.2-codex-high (Cursor Cloud originalModelName from run-info): investigation, local repro, patch, verification, and PR preparation.

中文

摘要

用已合入 main（#3443）的最小修复清除 InferenceX #3379 的 CPU Tests 失败。

失败并非 B300 native SRT power 覆盖改动本身引起，而是 #3410 将 clone 改为 git -c advice.detachedHead=false clone ... 后，测试夹具仍只匹配 $1 == clone，导致无法 staging checkout/。本草稿修复 PR 指向已有分支 fix/qwen35-b300-powerx-coverage，不接管 #3379 所有权。

验证

修复前 2 个代表用例失败；修复后整组 test_pool_launcher_stages_artifacts_and_propagates_failure 共 26 项通过。未使用 GPU，未为诊断重跑远程 CI。

与 main 冲突

#3379 相对当前 main 仍冲突（含 #3525 目录迁移）。本 PR 不解决 rebase，只修复当前谱系上的 CPU 测试契约问题。

AI 模型披露

• gpt-5.2-codex-high：调查、本地复现、补丁、验证与 PR 准备。