Native image generation: keep non-ASCII prompt text intact on Windows

CHANGELOG.md

@@ -11,6 +11,12 @@ permanent public record of what shipped and are never rewritten; the in-progress
 
 ## [Unreleased]
 
+### Fixed
+
+- **Image prompts with letters such as ß work with the native image backend on Windows.** A prompt
+  containing ß failed with "failed to encode prompt", and letters such as ä, ö and ü reached the
+  model garbled; the prompt now reaches stable-diffusion.cpp unchanged.
+
 ## [0.3.0] - 2026-10-14
 
 ### Added

localm/media/sdcpp/_child.py

@@ -110,6 +110,17 @@ def _prepare_environment(runtime_dir: Path) -> None:
         os.environ.pop(_HOST_VIS_VAR, None)
 
 
+def _use_c_ctype_locale() -> None:
+    """On Windows, set this process's C runtime LC_CTYPE to "C", so the byte-wise
+    tolower in stable-diffusion.cpp's CLIP tokenizer leaves UTF-8 prompt bytes
+    unchanged. Other platforms are left as they are.
+    See test_the_worker_keeps_utf8_prompt_bytes_through_tolower."""
+    if sys.platform != "win32":
+        return
+    import locale
+    locale.setlocale(locale.LC_CTYPE, "C")
+
+
 def _set_rocblas_tensile(extra_dirs) -> None:
     if os.environ.get("ROCBLAS_TENSILE_LIBPATH"):
         return
@@ -648,6 +659,7 @@ def worker_main(req_q, resp_q, cancel_event, crash_trace_path=None) -> None:
     install_parent_death_watchdog()
     ignore_interrupt_signals()
     suppress_native_error_dialogs()
+    _use_c_ctype_locale()
 
     log_cb, progress_cb, _State.preview_cb = _make_callbacks(resp_q)
     callbacks_set = False

tests/test_sdcpp_child_locale.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+"""The stable-diffusion.cpp worker keeps non-ASCII prompt bytes intact through
+the native tokenizer's byte-wise tolower."""
+
+import subprocess
+import sys
+import textwrap
+
+import pytest
+
+from localm.media.sdcpp import _child
+
+_PROMPT = "Ein weißer Fuchs läuft über einen vereisten See"
+
+_CRT_PROBE = textwrap.dedent("""
+    import ctypes, locale, sys
+    from localm.media.sdcpp import _child
+    crt = ctypes.CDLL("ucrtbase")
+    raw = sys.argv[1].encode("utf-8")
+
+    def lowered():
+        return bytes(crt.tolower(b) & 0xFF for b in raw)
+
+    locale.setlocale(locale.LC_CTYPE, ".1252")
+    before = lowered()
+    _child._use_c_ctype_locale()
+    after = lowered()
+    print(before.hex())
+    print(after.hex())
+""")
+
+
+@pytest.mark.skipif(sys.platform != "win32", reason="the code-page C runtime locale is Windows-only")
+def test_the_worker_keeps_utf8_prompt_bytes_through_tolower():
+    out = subprocess.run([sys.executable, "-c", _CRT_PROBE, _PROMPT],
+                         capture_output=True, text=True, timeout=60, check=True)
+    before_hex, after_hex = out.stdout.split()
+    before, after = bytes.fromhex(before_hex), bytes.fromhex(after_hex)
+    with pytest.raises(UnicodeDecodeError):
+        before.decode("utf-8")
+    assert after.decode("utf-8") == _PROMPT.lower()
+
+
+def test_the_worker_entry_point_sets_the_locale_before_serving(monkeypatch):
+    calls = []
+    monkeypatch.setattr(_child, "_use_c_ctype_locale", lambda: calls.append("locale"))
+    monkeypatch.setattr(_child, "_arm_crash_trace", lambda path: None)
+    import localm._mp_spawn as spawn
+    import localm.debuglog as debuglog
+    for name in ("install_parent_death_watchdog", "ignore_interrupt_signals",
+                 "suppress_native_error_dialogs"):
+        monkeypatch.setattr(spawn, name, lambda: None)
+    monkeypatch.setattr(debuglog, "attach_child_logging", lambda: None)
+
+    class _Queue:
+        def get(self):
+            calls.append("get")
+            return None
+
+    import threading
+    _child.worker_main(_Queue(), _Queue(), threading.Event())
+    assert calls[:2] == ["locale", "get"]
+
+
+def test_other_platforms_leave_the_locale_alone(monkeypatch):
+    import locale
+    seen = []
+    monkeypatch.setattr(sys, "platform", "linux")
+    monkeypatch.setattr(locale, "setlocale", lambda *a: seen.append(a))
+    _child._use_c_ctype_locale()
+    assert seen == []