diff --git a/.github/workflows/pytest-matrix.yml b/.github/workflows/pytest-matrix.yml
index e21e05b..6cf38cb 100644
--- a/.github/workflows/pytest-matrix.yml
+++ b/.github/workflows/pytest-matrix.yml
@@ -36,42 +36,15 @@ jobs:
       - uses: astral-sh/setup-uv@v7
         with:
           cache-dependency-glob: pyproject.toml
-        env:
-          # PER-RUNNER TOOL CACHE. Every runner on the Spartan lease inherited
-          # RUNNER_TOOL_CACHE=~/.runner-toolcache from the supervisor, so ~80
-          # runners shared ONE mutable cache on GPFS. @actions/tool-cache's
-          # writer is destructive, not additive:
-          #
-          #   _createToolPath():  rmRF(<ver>/<arch>)      <- deletes the tree
-          #                       rmRF(<ver>/<arch>.complete)
-          #                       mkdirP(<ver>/<arch>)
-          #   ...copy files in...
-          #   _completeToolPath(): write <arch>.complete
-          #
-          # and the reader, find(), treats dir + `.complete` as a HIT and puts
-          # it on PATH WITHOUT checking that the binary is inside. So when a new
-          # uv release makes every concurrent job a cache MISS at once, each one
-          # deletes the tree the others are using or filling. Two observed faces,
-          # same cause:
-          #   ENOENT copying into the toolcache  (two writers collide)
-          #   "Added .../uv/<ver>/x86_64 to the path" then
-          #   ".../uv: No such file or directory" -> exit 127
-          #     (a reader's hit gets deleted under it — figrecipe main, run
-          #      29945387939, 2026-07-22: all three matrix legs, three runners)
-          #
-          # Scoping the cache per runner removes the sharing, so no job can
-          # destroy another's tree. This is why PRE-SEEDING the image does NOT
-          # fix it: _createToolPath rmRF's a pre-seeded tree exactly the same.
-          #
-          # /tmp is node-local and may be cleared: that degrades to a cache MISS
-          # (re-download, slower), never a dangling path — mkdirP recreates it.
-          # Deliberately NOT on GPFS: punim0264 is at 98.2% of its 8M inode
-          # limit and a per-runner cache there would multiply inode use by ~80.
-          #
-          # Set at STEP level on purpose: the `runner` context is not available
-          # to job-level `env`. setup-uv writes the resolved path to $GITHUB_PATH,
-          # so later steps resolve uv without needing this variable themselves.
-          RUNNER_TOOL_CACHE: /tmp/scitex-ci-runner-work/toolcache/${{ runner.name }}
+        # NOTE: a per-runner RUNNER_TOOL_CACHE override was tried here and
+        # REMOVED as a no-op. RUNNER_TOOL_CACHE is a runner-reserved default;
+        # a workflow-level `env:` does NOT override it (measured 2026-07-23,
+        # scitex-hpc ci run 30028631599: the step still resolved the runner's
+        # ~/.runner-toolcache, not the override). The shared-toolcache flake
+        # (setup-uv concurrent-populate mutual destruction) can only be fixed
+        # where the RUNNER exports the variable — the launcher script
+        # (relaunch-scitex-hpc-runner.sh), per-runner + node-local. Tracked in
+        # scitex-hpc card ci-runner-toolcache-per-runner-launcher-ssot-20260723.
       - name: Install (.[all,dev]) + pytest
         run: |
           set -euo pipefail