ci: run hw-coord on per-node runners; source Vivado settings64.sh

Previously hw-coord ran on the coordinator host (lbvm), but lbvm
doesn't have Vivado installed.  XSA-pipeline tests that depend on
`sdtgen` therefore pytest-skipped silently — the matrix looked
green but tests weren't actually exercising any XSA logic on
bq/mini2.

Relocate hw-coord to the per-node self-hosted runners (same
`hw-<place>` label as hw-direct), where Vivado 2025.1 is available
at /tools/Xilinx/2025.1/.  Source settings64.sh before pytest in
both hw-direct and hw-coord so sdtgen / xsct / xsdb are on PATH
regardless of the runner service's login-shell setup.  Use
labgrid-client from the per-node adidt-venv (labgrid is a
transitive dep) so we don't need a separate labgrid-venv on each
runner.

The hw-coordinator runner now only carries the preflight job.

Author: tfcollins

.github/workflows/hardware-test.yml

@@ -140,6 +140,10 @@ jobs:
         if: steps.check_direct_env.outputs.skip != 'true'
         run: |
           set -euo pipefail
+          if [ -f /tools/Xilinx/2025.1/Vivado/settings64.sh ]; then
+            # shellcheck disable=SC1091
+            source /tools/Xilinx/2025.1/Vivado/settings64.sh
+          fi
           export LG_ENV="$LG_DIRECT_ENV"
           "$HOME/.cache/adidt-ci/adidt-venv/bin/pytest" \
             -p no:genalyzer -v -s ${{ join(matrix.node.tests, ' ') }}
@@ -152,7 +156,11 @@ jobs:
       matrix:
         node: ${{ fromJSON(needs.preflight.outputs.available_nodes) }}
     name: hw-coord (${{ matrix.node.place }})
-    runs-on: [self-hosted, hw-coordinator]
+    # Run on the per-node runner (same label as hw-direct) so the
+    # Xilinx toolchain (sdtgen, xsct) available there can drive the
+    # XSA pipeline.  The coordinator host (lbvm) lacks Vivado, so
+    # running pytest there silently skips every XSA-dependent test.
+    runs-on: [self-hosted, "${{ matrix.node.runner_label }}"]
     timeout-minutes: 40
     concurrency:
       group: hw-coord-${{ matrix.node.place }}
@@ -180,7 +188,9 @@ jobs:
           LG_COORDINATOR: ${{ env.COORDINATOR }}
         run: |
           set -euo pipefail
-          LGCLIENT="$HOME/.cache/adidt-ci/labgrid-venv/bin/labgrid-client"
+          # labgrid-client is a transitive dep of adidt's dev extras,
+          # so it lands in the per-node adidt-venv.
+          LGCLIENT="$HOME/.cache/adidt-ci/adidt-venv/bin/labgrid-client"
           "$LGCLIENT" -x "$LG_COORDINATOR" -p "${{ matrix.node.place }}" acquire
 
       - name: Run coordinator-mode tests
@@ -189,6 +199,12 @@ jobs:
           LG_ENV: ${{ github.workspace }}/${{ matrix.node.env_remote }}
         run: |
           set -euo pipefail
+          # Source Vivado settings so sdtgen / xsct / xsdb are on PATH
+          # for XSA-pipeline tests.
+          if [ -f /tools/Xilinx/2025.1/Vivado/settings64.sh ]; then
+            # shellcheck disable=SC1091
+            source /tools/Xilinx/2025.1/Vivado/settings64.sh
+          fi
           "$HOME/.cache/adidt-ci/adidt-venv/bin/pytest" \
             -p no:genalyzer -v -s ${{ join(matrix.node.tests, ' ') }}
 
@@ -198,5 +214,5 @@ jobs:
           LG_COORDINATOR: ${{ env.COORDINATOR }}
         run: |
           set -euo pipefail
-          LGCLIENT="$HOME/.cache/adidt-ci/labgrid-venv/bin/labgrid-client"
+          LGCLIENT="$HOME/.cache/adidt-ci/adidt-venv/bin/labgrid-client"
           "$LGCLIENT" -x "$LG_COORDINATOR" -p "${{ matrix.node.place }}" release || true

doc/source/developer/hardware_ci.rst

@@ -28,9 +28,15 @@ The workflow has three jobs:
   ``LG_DIRECT_ENV`` is unset on a runner the job skips all work and
   exits green — add the env entry in ``~/actions-runner/.env`` to
   light up direct-mode on that node.
-- **hw-coord** — a matrix job, one leg per available node, all
-  running on the single ``hw-coordinator`` runner.  Each leg uses the
-  committed ``env_remote_<place>.yaml`` that the manifest declares.
+- **hw-coord** — a matrix job, one leg per available node, running
+  on the **same per-node runner as hw-direct** (label
+  ``hw-<place>``).  Each leg uses the committed
+  ``env_remote_<place>.yaml`` from the manifest and talks to the
+  labgrid coordinator via ``LG_COORDINATOR`` — so pytest exercises
+  the coordinator code path while the XSA toolchain (Vivado's
+  sdtgen, xsct) and kernel build artifacts already present on the
+  per-node runner are used in place.  The ``hw-coordinator`` runner
+  itself carries only the preflight job.
 
 Fork-PR protection is handled by GitHub Actions' built-in workflow
 approval gate rather than a custom ``environment:``.  Under *Settings