ci: parallel test shards, nested unit tests, and a pre-push GIT_DIR leak (draft, measuring)

.githooks/pre-push

@@ -54,6 +54,26 @@ def _git(*args: str) -> subprocess.CompletedProcess:
     )
 
 
+def _drop_repo_env() -> None:
+    """Remove the variables that bind a git process to THIS repository.
+
+    git exports GIT_DIR to hooks (in a linked worktree it is always set), and
+    every layer below inherits os.environ. The review lenses are agents that
+    may run the test suite; a test that does `git init` / `add -A` / `commit`
+    in a temp dir then writes to the pushing repo instead. On 2026-10-29 one
+    push left 5 fixture commits on the branch (the tree reduced to 8 files) and
+    `[user] name = t` in the shared .git/config. The hook has already chdir'd
+    to the worktree root, so git rediscovers the same repo without them.
+    `--local-env-vars` is git's own list: what it clears for a submodule.
+    """
+    names = _git("rev-parse", "--local-env-vars").stdout.split() or [
+        "GIT_DIR", "GIT_WORK_TREE", "GIT_INDEX_FILE", "GIT_COMMON_DIR",
+        "GIT_OBJECT_DIRECTORY", "GIT_PREFIX",
+    ]
+    for name in names:
+        os.environ.pop(name, None)
+
+
 def _source_local_env(repo_root: Path) -> None:
     """Parse simple KEY=VALUE lines from $MINI_ORK_HOME/config/local.sh.
 
@@ -176,6 +196,8 @@ def main() -> int:
     except OSError:
         return 0
 
+    _drop_repo_env()
+
     # Per-developer local env preferences (non-secret).
     _source_local_env(REPO_ROOT)
 

.github/workflows/ci.yml

@@ -58,18 +58,24 @@ jobs:
   python-tests:
     name: python tests (${{ matrix.python-version }} / ${{ matrix.shard }})
     runs-on: ubuntu-latest
-    # The unit shards run 22-33 min and drift with every added test file (the
-    # odd/even split reshuffles); 35 cancelled a passing 3.12/unit-a at 35.4
-    # min on 2026-10-28 (a re-run took 28.6). Headroom, not a ceiling to aim at.
-    timeout-minutes: 50
+    # Each shard runs its tests on every runner core (pytest-xdist, below).
+    # Serially the unit shards took 23-39 min and 50 was the timeout; a shard
+    # that now needs more than 30 is a hang or a regression worth seeing.
+    timeout-minutes: 30
     strategy:
       fail-fast: false
       matrix:
         python-version: ["3.11", "3.12"]
-        # Sharding: the suite is ~2,090 tests dominated by bash<->python
-        # parity subprocess costs (~14-17 min as one job). Split the big
-        # tests/unit tier odd/even by file (interleave balances the slow
-        # clusters) and run everything else as a third shard.
+        # Sharding: tests/unit is ~14,000 tests whose time is spread thin —
+        # subprocess and git work in hundreds of files, the slowest file ~4%
+        # of the total — so parallelism, not trimming single tests, is what
+        # moves it. Split it odd/even by file and run everything else as a
+        # third shard; within a shard, xdist uses all cores (`-n auto`). Both
+        # tiers pass on 8 workers (unit: 14,000 passed, 0 failed; rest: 734
+        # passed, its one local failure reproduces serially; 2026-10-29), so
+        # tests do not share state across processes.
+        # `find`, not `ls tests/unit/*.py`: the glob never reached
+        # tests/unit/acp/, so its 28 tests ran in no shard at all.
         shard: [unit-a, unit-b, rest]
     steps:
       - uses: actions/checkout@v7
@@ -90,7 +96,7 @@ jobs:
           # tests/unit/test_acp_agent_py.py imports at module scope; without
           # the extra the whole shard dies at collection, not on an assertion.
           pip install -e ".[yaml,acp]"
-          pip install pytest fastapi uvicorn httpx2
+          pip install pytest "pytest-xdist>=3.5" fastapi uvicorn httpx2
 
       # The web smoke suite points the FastAPI route handlers at a real
       # state.db; init bootstraps the full schema so those tests run
@@ -100,15 +106,15 @@ jobs:
 
       - name: Run pytest (unit shard A)
         if: matrix.shard == 'unit-a'
-        run: ls tests/unit/*.py | awk 'NR % 2 == 1' | xargs python -m pytest -q -p no:cacheprovider
+        run: find tests/unit -name 'test_*.py' | sort | awk 'NR % 2 == 1' | xargs python -m pytest -q -p no:cacheprovider -n auto
 
       - name: Run pytest (unit shard B)
         if: matrix.shard == 'unit-b'
-        run: ls tests/unit/*.py | awk 'NR % 2 == 0' | xargs python -m pytest -q -p no:cacheprovider
+        run: find tests/unit -name 'test_*.py' | sort | awk 'NR % 2 == 0' | xargs python -m pytest -q -p no:cacheprovider -n auto
 
       - name: Run pytest (root + non-unit layers)
         if: matrix.shard == 'rest'
-        run: python -m pytest tests/ -q -p no:cacheprovider --ignore=tests/unit
+        run: python -m pytest tests/ -q -p no:cacheprovider -n auto --ignore=tests/unit
 
   # The optional `runtime` extra pins `verifiers>=0.2.0,<0.5`, and that ceiling moves
   # by dependabot PR (it went <0.3 -> <0.5 on 2026-10-27 with no check on the API).