fix(benchmarks): two generated graders reach the same verdict off POSIX

Both have failed every run of CI's windows-latest job since the corpus
landed, and neither is a flake.

gen-topo-cycle timed each call with signal.alarm. Windows Python has no
SIGALRM, so the grader raised while installing the handler, before it
graded anything. A daemon worker thread keeps the semantics exactly —
_Timeout on a hang, the original exception re-raised, so the cyclic case
still reads ValueError — and cannot hold the interpreter open at exit.

gen-csv-quotes compared cli.py's stdout against an expected string.
Windows Python writes CRLF in text mode, so that comparison could never
hold; the carriage returns come off first, because the grader checks the
fields and not the host's line endings. Its mktemp template also read
TMPDIR, which under Git Bash is a Windows path — the rest of the corpus
asks for a bare mktemp -d, and now so does this one.

benchmarks/train/tasks/gen-csv-quotes/verify.sh

@@ -4,7 +4,7 @@
 # so it references csv_reader.py and cli.py directly.
 set -e
 
-tmpdir="$(mktemp -d "${TMPDIR:-/tmp}/csv-verify.XXXXXX")"
+tmpdir="$(mktemp -d)"
 trap 'rm -rf "$tmpdir"' EXIT
 
 # 1. Behavioural checks against the public API (python3 stdlib only).
@@ -46,7 +46,9 @@ name,note
 EOF
 
 expected=$'name|note\nAda, Countess of Lovelace|wrote "notes"'
-got="$(python3 cli.py "$tmpdir/sample.csv")"
+# Windows' python writes CRLF in text mode, so strip the carriage returns
+# before comparing: the grader checks the fields, not the host's line endings.
+got="$(python3 cli.py "$tmpdir/sample.csv" | tr -d '\r')"
 if [ "$got" != "$expected" ]; then
     printf 'FAIL: cli.py printed:\n%s\n\nexpected:\n%s\n' "$got" "$expected" >&2
     exit 1

benchmarks/train/tasks/gen-topo-cycle/verify.sh

@@ -5,8 +5,8 @@ set -e
 cd "$(dirname "$0")"
 
 python3 - <<'PYEOF'
-import signal
 import sys
+import threading
 
 import topo
 
@@ -15,19 +15,26 @@ class _Timeout(Exception):
     pass
 
 
-def _alarm_handler(signum, frame):
-    raise _Timeout("topological_sort did not return within the time budget")
-
-
-signal.signal(signal.SIGALRM, _alarm_handler)
-
-
+# A worker thread rather than signal.alarm: Windows has no SIGALRM, and this
+# grader has to reach the same verdict on every host. The thread is a daemon so
+# a solution that never returns cannot hold the interpreter open at exit.
 def run_with_timeout(fn, *args, **kwargs):
-    signal.alarm(5)
-    try:
-        return fn(*args, **kwargs)
-    finally:
-        signal.alarm(0)
+    outcome = {}
+
+    def call():
+        try:
+            outcome["value"] = fn(*args, **kwargs)
+        except BaseException as exc:
+            outcome["error"] = exc
+
+    worker = threading.Thread(target=call, daemon=True)
+    worker.start()
+    worker.join(5)
+    if worker.is_alive():
+        raise _Timeout("topological_sort did not return within the time budget")
+    if "error" in outcome:
+        raise outcome["error"]
+    return outcome["value"]
 
 
 def is_topo_order(n, edges, order):