fix(e2e): the Windows job's locale, not the guard, decided what 233 could read

`ci-windows-e2e` went red on 233 with

    UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position 3037

from §7, whose `read_text()` names no encoding. Python then decodes with the
LOCALE encoding — UTF-8 on the Linux/macOS runners, cp1252 on the Windows one —
and every file this test reads (matrix.json, the engine adapters, the READMEs)
contains non-ASCII.

The crash is the benign half. cp1252 leaves only five byte values undefined, so
the reads that DON'T hit one of them succeed and return mojibake: the regex
matches nothing and the guard prints its success line while checking nothing.
§1 has been reading matrix.json that way from the start and stayed green purely
because its non-ASCII bytes missed those five. Same defect, opposite symptom —
and the silent one is the symptom this whole test exists to catch, so it must
not be the test's own failure mode.

Every read now names utf-8, including the `subprocess(text=True)` that decoded
git's stdout the same way. And the test exports PYTHONWARNDEFAULTENCODING /
PYTHONWARNINGS=error::EncodingWarning so an unspecified encoding is a hard error
on the FIRST machine that runs it — this class of bug should not be discoverable
only on Windows.

Verified both directions: with §7's encoding removed the test exits 1 locally
(EncodingWarning), restored it exits 0. The injection asserts its own anchor
first, because a sabotage that silently fails to apply reads exactly like a
passing test.

Author: speak-agent

tests/e2e/233_bench_matrix.sh

@@ -19,6 +19,30 @@
 #   4. the workflow reads the file instead of repeating it.
 set -e
 
+# ⚠️ EVERY python read below MUST name its encoding, and this is what enforces it.
+#
+# `open()`, `read_text()` and `subprocess(text=True)` decode with the LOCALE
+# encoding, which is UTF-8 on the Linux and macOS runners and cp1252 on the
+# Windows one. Every file this test reads — matrix.json, the engine adapters,
+# the READMEs — contains non-ASCII, so on Windows the reads either raise
+#
+#     UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position 3037
+#
+# or, for the bytes cp1252 does happen to map, silently produce mojibake: the
+# regex then matches nothing and the guard reports success while guarding
+# nothing. That is the same "failure looks like success" shape this whole test
+# exists to catch, so it must not be the test's own failure mode.
+#
+# §1 read matrix.json without an encoding for a while and stayed green purely
+# because its non-ASCII bytes missed cp1252's five undefined ones; §7 hit 0x8f
+# and turned the whole Windows e2e job red. Both are the same defect.
+#
+# These two variables turn an unspecified encoding into a hard error, so the
+# next one fails on the FIRST machine that runs it rather than only on Windows.
+# Ignored by Python < 3.10, which predates EncodingWarning.
+export PYTHONWARNDEFAULTENCODING=1
+export PYTHONWARNINGS=error::EncodingWarning
+
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
 MATRIX="$ROOT/bench/matrix.json"
 WORKFLOW="$ROOT/.github/workflows/bench.yml"
@@ -34,7 +58,7 @@ SPEC="$ROOT/bench/SPEC.md"
 python3 - "$MATRIX" "$ROOT" <<'PY'
 import json, os, re, sys
 
-m = json.load(open(sys.argv[1]))
+m = json.load(open(sys.argv[1], encoding="utf-8"))
 axes = m["axes"]
 fail = []
 
@@ -186,7 +210,11 @@ tracked = subprocess.run(
     ["git", "-C", root, "ls-files",
      "bench/projects/*/.xmake*", "bench/projects/*/build/*",
      "bench/projects/*/CMakeCache.txt", "bench/projects/*/bazel-*"],
-    capture_output=True, text=True).stdout.split()
+    # encoding pinned, not `text=True` alone: that decodes the child's stdout
+    # with the LOCALE encoding, which on a Windows runner is cp1252. A path (or
+    # any UTF-8 byte) then either raises or, worse, mojibakes into something
+    # that no longer matches — a guard that silently stops guarding.
+    capture_output=True, encoding="utf-8").stdout.split()
 if tracked:
     fail.append("engine scratch is tracked in git (machine-local state, and one of "
                 f"these froze a fixed bug into CI): {tracked[:4]}"
@@ -211,7 +239,7 @@ if not re.match(r"^\d+(\.\d+)+$", str(m.get("reference_mcpp", ""))):
 # some other release, with every ratio still looking perfectly reasonable.
 xlings_pin = os.path.join(root, ".xlings.json")
 if os.path.isfile(xlings_pin):
-    ws = json.load(open(xlings_pin)).get("workspace", {}).get("mcpp")
+    ws = json.load(open(xlings_pin, encoding="utf-8")).get("workspace", {}).get("mcpp")
     if ws and ws != m.get("reference_mcpp"):
         fail.append(f"reference_mcpp={m.get('reference_mcpp')} but .xlings.json bootstraps "
                     f"mcpp {ws} — the reference arm IS the bootstrapped binary, so these "
@@ -239,7 +267,7 @@ PY
 python3 - "$MATRIX" "$ROOT/bench/src/spec.cppm" "$ROOT/bench/src/registry.cppm" <<'PY'
 import json, re, sys
 
-m = json.load(open(sys.argv[1]))
+m = json.load(open(sys.argv[1], encoding="utf-8"))
 spec = open(sys.argv[2], encoding="utf-8").read()
 registry = open(sys.argv[3], encoding="utf-8").read()
 fail = []
@@ -283,7 +311,7 @@ if not os.path.isfile(data):
     raise SystemExit(0)
 
 truth = {}
-for c in json.load(open(data))["cells"]:
+for c in json.load(open(data, encoding="utf-8"))["cells"]:
     if c["status"] == "ok":
         truth.setdefault(c["engine"], {})[c["scenario"]] = round(c["median_s"], 2)
 default = next((k for k in truth if k.startswith("mcpp@") and "+" not in k), None)
@@ -332,7 +360,7 @@ if grep -qE '^\s*case ",\$want," in \*,(linux|macos|windows),\*\)' "$WORKFLOW";
     echo "FAIL: bench.yml still enumerates platforms inline; matrix.json owns that list"
     exit 1
 fi
-for img in $(python3 -c "import json,sys;print(' '.join(json.load(open(sys.argv[1]))['runners'].values()))" "$MATRIX"); do
+for img in $(python3 -c "import json,sys;print(' '.join(json.load(open(sys.argv[1], encoding='utf-8'))['runners'].values()))" "$MATRIX"); do
     if grep -q "runs-on: $img" "$WORKFLOW"; then
         echo "FAIL: bench.yml hard-codes runner image '$img'; it must come from matrix.json"
         exit 1
@@ -359,7 +387,7 @@ grep -q 'matrix.json' "$SPEC" \
 python3 - "$ROOT" <<'PY' || exit 1
 import json, pathlib, re, sys
 root = pathlib.Path(sys.argv[1])
-m = json.loads((root / "bench/matrix.json").read_text())
+m = json.loads((root / "bench/matrix.json").read_text(encoding="utf-8"))
 bad = []
 for c in m["cells"]:
     proj = c.get("project", "")
@@ -376,7 +404,7 @@ for c in m["cells"]:
         # failed exactly that way.
         if not f.exists():
             continue
-        body = re.sub(r"#.*", "", f.read_text())
+        body = re.sub(r"#.*", "", f.read_text(encoding="utf-8"))
         # ANY rule, not `cc_*` specifically. Every bazel rule instantiation
         # carries a `name =` attribute; `load()`, `package()` and
         # `exports_files()` do not. Matching `cc_binary|cc_library` was wrong:
@@ -420,7 +448,7 @@ for f in root.glob("*.cppm"):
     # sees the resolved path.
     if f.name == "engine.cppm":
         continue
-    for n, line in enumerate(f.read_text().splitlines(), 1):
+    for n, line in enumerate(f.read_text(encoding="utf-8").splitlines(), 1):
         code = line.split("//", 1)[0]
         if re.search(r'compiler\s*==\s*"(clang|gcc)"', code):
             bad.append(f"{f.name}:{n}: {line.strip()[:90]}")