[DO NOT MERGE] Debug Proxy

.buildkite/windows.rayci.yml

@@ -18,7 +18,8 @@ steps:
         --operating-system windows
         --except-tags no_windows
         --test-env=CI="1"
-        --test-env=USERPROFILE
+        --test-env=USERPROFILE || true
+      - sleep 28800
     depends_on: windowsbuild
 
   # block full test suite on premerge and microcheck

ci/docker/windows.build.Dockerfile

@@ -5,5 +5,17 @@ ENV PYTHON=3.10
 ENV PYTHON_FULL_VERSION=3.10.19
 ENV RAY_BUILD_ENV=win_py$PYTHON_FULL_VERSION
 
+# build_base.sh builds the test interpreter in a dedicated conda env ("ray")
+# rather than swapping the base env's python in place.
+ENV RAY_CONDA_ENV=ray
+
 COPY . .
 RUN bash ci/ray_ci/windows/build_base.sh
+
+# Make the dedicated env the default interpreter for build_ray.sh and for bazel
+# at test time. On this base image the effective PATH lives in the Windows
+# registry (machine scope), NOT in a Docker ENV -- so we prepend the env there.
+# Using `ENV PATH=...;${PATH}` would expand ${PATH} to nothing and wipe
+# conda/git/bash off the PATH (CreateProcess: file not found). `$env:Path` below
+# is expanded by powershell at build time (Docker does not substitute in RUN).
+RUN powershell -Command "[Environment]::SetEnvironmentVariable('Path', 'C:\Miniconda3\envs\ray;C:\Miniconda3\envs\ray\Library\mingw-w64\bin;C:\Miniconda3\envs\ray\Library\usr\bin;C:\Miniconda3\envs\ray\Library\bin;C:\Miniconda3\envs\ray\Scripts;C:\Miniconda3\envs\ray\bin;' + $env:Path, 'Machine')"

ci/ray_ci/windows/build_base.sh

@@ -12,13 +12,31 @@ powershell ci/ray_ci/windows/install_bazelisk.ps1
 powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/0.9.22/install.ps1 | iex"
 
 conda init
-# TODO(ci): Remove once conda fixes the splitext bug in delete.py (conda/conda#15760).
-# Conda 26.3.1 crashes on Windows when it tries to clean up a locked exe during
-# a build-variant swap. Preventing self-update avoids that code path.
-conda config --set auto_update_conda false
-conda install -q -y python="${PYTHON_FULL_VERSION}" requests=2.32.3
+
+# Build the test interpreter in a dedicated conda env instead of swapping the
+# base env's python in place. The base image ships Python 3.8 in `base`, and an
+# in-place `conda install python=${PYTHON_FULL_VERSION}` either crashes
+# (conda#15760) or -- with auto_update_conda disabled -- silently no-ops and
+# leaves `base` on 3.8, which then mismatches the py310 bazel deps and blows up
+# at test time. A fresh env sidesteps both failure modes. We put it first on
+# PATH below for this build; windows.build.Dockerfile does the same on the
+# image's machine PATH so bazel resolves to it at test time.
+conda create -y -n "${RAY_CONDA_ENV}" python="${PYTHON_FULL_VERSION}" requests=2.32.3
 # Force CA trust stack to the newest versions available at build time.
-conda update -c conda-forge -q -y ca-certificates certifi
+conda update -n "${RAY_CONDA_ENV}" -c conda-forge -y ca-certificates certifi
+
+# Put the env first on PATH for the rest of this build so `python`/`pip` below
+# resolve to it. (Test-time PATH is set on the image's machine PATH by
+# windows.build.Dockerfile, because on Windows the base PATH lives in the
+# registry, not in Docker ENV.)
+env_root="/c/Miniconda3/envs/${RAY_CONDA_ENV}"
+export PATH="${env_root}:${env_root}/Library/mingw-w64/bin:${env_root}/Library/usr/bin:${env_root}/Library/bin:${env_root}/Scripts:${env_root}/bin:${PATH}"
+
+# Fail the build loudly here if the interpreter on PATH is not the expected
+# version, rather than letting the mismatch surface as a confusing import error
+# (e.g. `module 'typing' has no attribute '_SpecialGenericAlias'`) at test time.
+python --version
+python -c "import sys; v = '%d.%d' % sys.version_info[:2]; exp = '${PYTHON}'; assert v == exp, f'expected python {exp} on PATH, got {sys.version}'"
 
 # Install torch first, as some dependencies (e.g. torch-spline-conv) need torch to be
 # installed for their own install.