Run the Python tests on six shards

.github/workflows/pr_code_changes.yaml

@@ -219,7 +219,7 @@ jobs:
         uses: ./.github/actions/test-environment
       - name: Test documentation builds
         run: make documentation
-  # The YAML tests, the Python tests on five runners, and the check that
+  # The YAML tests, the Python tests on six runners, and the check that
   # every test ran once and passed, in the locked environment. As one job
   # they took 59 minutes on 10 October 2026 (run 38075384588), and runs were
   # cancelled at a 60-minute limit (issues #2099 and #2257). The release

.github/workflows/test-suite.yaml

@@ -4,10 +4,10 @@
 # publish. All three run the same jobs in the same locked environment.
 #
 # Jobs: the YAML policy tests (which also list every Python test), the Python
-# tests split over five runners, and Result, which fails unless every listed
+# tests split over six runners, and Result, which fails unless every listed
 # Python test ran in exactly one shard and passed, apart from the outcomes
 # accepted in .github/test_outcome_allowlist.toml. A caller that names its job
-# Test gets the checks "Test / YAML", "Test / Python K/5" and "Test / Result";
+# Test gets the checks "Test / YAML", "Test / Python K/6" and "Test / Result";
 # "Test / Result" is the one to require.
 #
 # Runners. These jobs, and the check jobs that pull requests, merge queue
@@ -153,6 +153,16 @@ jobs:
     # request adding about a fifth more parameters (#1855) hit the timeout
     # on shard 1/4 twice, and its other shards took 48 to 59 minutes (run
     # 39568525017).
+    #
+    # Six, not five, for both. On 26 October a pull request's shards took
+    # 63 to 75 minutes, and the 75-minute limit cancelled shard 4/5 (run
+    # 40308966462, first attempt); a merge queue run that evening took 66
+    # to 69 (run 40406434741). The same evening one pull request lost
+    # shards 2/5, 3/5 and 5/5 to memory at 15.6 to 15.9 GB (run
+    # 40393666784), and the merge queue run for #2105 lost shard 2/5 at
+    # 15.5 GB (run 40395541241). With six, the listed slow tests come to
+    # about 134 minutes of worker time per shard instead of 161, and each
+    # worker runs about a sixth fewer tests.
     name: Python ${{ matrix.shard }}/${{ strategy.job-total }}
     if: ${{ !inputs.skip }}
     runs-on: ${{ vars.CI_RUNNER_ARM != '' && (vars.CI_RUNNER_EVENTS == '' || contains(format(',{0},', vars.CI_RUNNER_EVENTS), format(',{0},', github.event_name))) && vars.CI_RUNNER_ARM || 'ubuntu-24.04-arm' }}
@@ -168,7 +178,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        shard: [1, 2, 3, 4, 5]
+        shard: [1, 2, 3, 4, 5, 6]
     steps:
       - name: Checkout repo
         uses: actions/checkout@v6

Makefile

@@ -24,7 +24,7 @@ test-yaml:
 
 # Every Python test, serially. To spread them over four cores add
 # PYTEST_ARGS="-n 4 --dist worksteal" (pytest-xdist); to run what one CI
-# shard runs, add "--shard K/5" (policyengine_uk/tests/conftest.py). CI runs
+# shard runs, add "--shard K/6" (policyengine_uk/tests/conftest.py). CI runs
 # these tests through .github/workflows/test-suite.yaml, in the locked
 # environment (uv sync --frozen --extra dev), and also checks each test's
 # outcome against .github/test_outcome_allowlist.toml.

changelog.d/ci-six-test-shards.changed.md

@@ -0,0 +1 @@
+Split the Python tests in CI over six runners instead of five.

docs/book/validation/ubuntu-vs-macos.md

@@ -103,7 +103,7 @@ That change verified, on main at `3c48247e`, that:
 ## Checking for a regression
 
 CI now runs the test suite (`.github/workflows/test-suite.yaml`: the
-YAML tests in one job and the Python tests split over five shards) on
+YAML tests in one job and the Python tests split over six shards) on
 Linux arm64 (`ubuntu-24.04-arm`), and the install and smoke-import checks
 on x86 `ubuntu-latest`. There is no macOS job. If an OS difference is
 suspected again:

docs/engineering/skills/testing.md

@@ -86,10 +86,10 @@ the same suite,
 tree; a release tests the wheel it is about to publish, installed on its own.
 
 The suite runs the YAML policy tests in one job and splits the Python tests
-over five runners. Each runner runs one shard:
+over six runners. Each runner runs one shard:
 
 ```bash
-pytest policyengine_uk/tests -n 4 --dist worksteal --shard K/5 -o xfail_strict=true
+pytest policyengine_uk/tests -n 4 --dist worksteal --shard K/6 -o xfail_strict=true
 ```
 
 `--shard K/N` is defined in `policyengine_uk/tests/conftest.py`. It keeps the