chore: drop remaining Chinese-locale mentions, ignore docs/research

.gitignore

@@ -113,3 +113,6 @@ _ui_backup/
 # The whole directory can be deleted at any time; deleting it just means
 # it gets re-downloaded next time.
 /.cache/
+
+# Local research notes (contain private training names)
+/docs/research/

CHANGELOG.md

@@ -1613,7 +1613,7 @@ Test image generation + prior generation + Setup rewrite + Settings split + CLTa
   - Backend: `versions.list_project_state_ckpts()` / `list_project_lora_ckpts()` + `/api/projects/{pid}/state_ckpts` / `/lora_ckpts` endpoints
   - Fixes the root UX problem: previously you had to dig five levels deep from REPO_ROOT to `output/training_state_step*.pt` just to resume
 - **Setup rewrite: GPU-aware torch on first install + venv stale check + --reinstall as a lifeline**
-  - `studio.bat` is pure ASCII now (no more crashes under cp936 cmd.exe) + a test as backstop
+  - `studio.bat` is pure ASCII now (no more crashes under non-UTF-8 cmd.exe) + a test as backstop
   - bootstrap: prefers `py -3` on Windows, iterates version checks on Linux
   - venv stale check + `--reinstall` flag (environment lifeline)
   - GPU-aware torch on first install; big warning when a CPU-only build gets installed by mistake

docs/announcements/2026-05-09-v0.5.0.md

@@ -26,7 +26,7 @@ Test image generation + prior generation + Setup rewrite + Settings split + CLTa
   - Backend: `versions.list_project_state_ckpts()` / `list_project_lora_ckpts()` + `/api/projects/{pid}/state_ckpts` / `/lora_ckpts` endpoints
   - Fixes the root UX problem: previously you had to dig five levels deep from REPO_ROOT to `output/training_state_step*.pt` just to resume
 - **Setup rewrite: GPU-aware torch on first install + venv stale check + --reinstall as a lifeline**
-  - `studio.bat` is pure ASCII now (no more crashes under cp936 cmd.exe) + a test as backstop
+  - `studio.bat` is pure ASCII now (no more crashes under non-UTF-8 cmd.exe) + a test as backstop
   - bootstrap: prefers `py -3` on Windows, iterates version checks on Linux
   - venv stale check + `--reinstall` flag (environment lifeline)
   - GPU-aware torch on first install; big warning when a CPU-only build gets installed by mistake

docs/design/preprocess-train-scope-plan.md

@@ -893,5 +893,5 @@ moves from project level to version/train/. Module boundaries are unchanged.
 - `studio/api/routers/projects/curation.py:117-186,271,342` — thumb endpoint bucket=train + 2 duplicates endpoint URLs
 - `studio/web/src/components/Sidebar.tsx:11-29,266-274,472` — STEP order / phase mapping / project-scope regex
 - `studio/web/src/api/client.ts` — `VersionPhase` / `PHASE_ORDER` / `PHASE_SKIPPABLE`
-- `studio/web/src/i18n/locales/{zh,en}.json` — "(optional)" copy + idx numbering
+- `studio/web/src/i18n/locales/{ru,en}.json` — "(optional)" copy + idx numbering
 - `studio/web/src/pages/project/steps/Preprocess*.tsx` — data contract + 4 sub-pages

docs/design/queue-pause-spike-report.md

@@ -79,7 +79,7 @@ was never emitted."
 ### Parent stdout's own encoding
 
 A minor pitfall unrelated to the main project: on Windows, the shell's default
-codepage (cp936 / cp932) raises `UnicodeEncodeError` when encoding non-ASCII
+ANSI codepage raises `UnicodeEncodeError` when encoding non-ASCII
 `print` output. `env["PYTHONIOENCODING"] = "utf-8"` only takes effect for the
 **child process** — the parent itself needs
 `sys.stdout.reconfigure(encoding="utf-8")`. The real supervisor writes to

runtime/anima_train.py

@@ -40,7 +40,7 @@
     if _ps not in sys.path:
         sys.path.insert(0, _ps)
 
-# The Windows console defaults to cp936; logging / print writing non-ASCII text
+# The Windows console defaults to the ANSI codepage; logging / print writing non-ASCII text
 # raises UnicodeEncodeError, and the default handler's errors='backslashreplace'
 # turns it into \uXXXX escapes -- that's the source of the garbled text seen in task logs.
 # Force stdout/stderr to UTF-8 + replace so non-ASCII text / emoji always print directly.

studio/domain/errors.py

@@ -20,11 +20,11 @@
   - the `message` field is **English** -- the frontend looks up `code` in an
     i18n table to render a localized string; `message` is only the English
     fallback display. This avoids reviving the ADR-0008 §cross-D
-    Chinese-string-matching trap through DomainError.
+    localized-string-matching trap through DomainError.
   - the `code` field is named **domain.action** (`preset.not_found` /
     `curation.duplicate`).
   - short-term (0.12.0), existing service errors like PresetError still have
-    Chinese messages (C3 only adds the base class, doesn't touch messages),
+    legacy messages (C3 only adds the base class, doesn't touch messages),
     but **new code must** use an English message + i18n code.
 
 Usage:

studio/infrastructure/logging.py

@@ -230,7 +230,7 @@ def __init__(self) -> None:
 
 
 def reconfigure_console_utf8() -> None:
-    """Windows consoles default to cp932/cp936, so writing non-ASCII text / emoji raises
+    """Windows consoles default to a legacy ANSI codepage, so writing non-ASCII text / emoji raises
     UnicodeEncodeError. Forces stdout/stderr into UTF-8 + replace mode so the logger never
     raises.
 

studio/services/runtime/flash_attention.py

@@ -106,7 +106,7 @@ def detect_env() -> dict[str, Any]:
     # the driver version is always stored separately as `driver_cuda_ver` for the UI to display and for troubleshooting (so the user can see
     # a case like "driver supports cu130, PyTorch is cu128" and immediately understand why the wheel should be cu128).
     driver_cuda_ver: Optional[str] = None
-    # Historically only caught (subprocess.SubprocessError, OSError); on a Windows Chinese locale (cp936),
+    # Historically only caught (subprocess.SubprocessError, OSError); on a non-UTF-8 Windows locale,
     # decoding nvidia-smi output with text=True could raise UnicodeDecodeError (not one of those subclasses),
     # causing a raw 500. Switched to errors='replace' to avoid that, and widened the except to Exception.
     try:

studio/supervisor/core.py

@@ -1365,7 +1365,7 @@ def _popen(
         creationflags = 0
         if os.name == "nt":
             creationflags = subprocess.CREATE_NEW_PROCESS_GROUP  # type: ignore[attr-defined]
-        # Windows defaults stdout to cp936; any worker writing Chinese text / emoji
+        # Windows defaults stdout to the ANSI codepage; any worker writing non-ASCII text / emoji
         # would trigger a UnicodeEncodeError, and logging's default backslashreplace
         # would turn it into \uXXXX, filling the task log with garbled text. Give
         # every child process a UTF-8 + unbuffered fallback here.

studio/web/src/api/client.ts

@@ -2180,16 +2180,6 @@ export interface ModelsRootMigrateStatus {
   error: string
 }
 
-export interface AnnouncementPost {
-  id: string
-  date: string
-  tag: 'release' | 'notice' | 'migration'
-  title: { zh: string; en: string }
-  body: { zh: string; en: string }
-  pin: boolean
-  version: string | null
-}
-
 export const api = {
   health: () => req<HealthResponse>('/api/health'),
   systemStats: () => req<SystemStats>('/api/system/stats'),

studio/web/src/components/TagsInput.tsx

@@ -13,7 +13,7 @@ export function parseTags(s: string): string[] {
  * - **Editing (focus)**: a plain-text `<input>`, commas / spaces typed freely. Controlled by the text, not the
  *   array, to avoid a "join-back on every keystroke" wiping out a comma / trailing space being typed. An autocomplete
  *   popover pops up below the input; keyboard up/down + Enter/Tab select.
- * - **Idle (blur)**: renders tags as chips, scannable at a glance; shows the Chinese translation when there's a match.
+ * - **Idle (blur)**: renders tags as chips, scannable at a glance.
  *
  * On blur the text is normalized to `tags.join(', ')`, so re-entering edit mode shows the canonical form.
  * Use this directly for cases that already have their own outer label (Settings' SettingsField); for a

studio/web/src/pages/project/steps/Train.tsx

@@ -441,7 +441,7 @@ export default function TrainPage() {
     } catch (e) {
       const msg = String(e)
       // Match on the error code, not on the message: the message is localized
-      // (and was never Chinese to begin with), so a text match silently turned
+      // so a text match silently turned
       // the overwrite prompt into a plain error toast.
       if ((e as ApiError).code === 'preset.exists') {
         const overwrite = await confirm(t('train.alreadyExists', { name: trimmed }), {

studio/web/src/pages/tools/Settings.tsx

@@ -252,7 +252,7 @@ const MODEL_DESCRIPTION_KEYS: Record<string, string> = {
   cltagger: 'settings.modelDescriptions.cltagger',
 }
 
-// Catalog names come from the backend in Chinese for the main models.
+// The backend sends English catalog names; localize the main models on the frontend.
 const MODEL_NAME_KEYS: Record<string, string> = {
   anima_main: 'settings.modelNames.animaMain',
   krea2_main: 'settings.modelNames.krea2Main',

studio/workers/_base.py

@@ -10,7 +10,7 @@
     - calling run + sys.exit
 
 Optional helper `reconfigure_console_utf8()` -- the Windows console defaults to
-cp932/cp936, so writing non-ASCII / emoji raises UnicodeEncodeError; calling this once
+a legacy ANSI codepage, so writing non-ASCII / emoji raises UnicodeEncodeError; calling this once
 switches stdout/stderr to UTF-8 + replace mode. Currently only tag_worker needs it (other
 workers don't write non-ASCII captions).
 """
@@ -68,7 +68,7 @@ def _worker_kind_from_argv() -> str:
 
 
 def reconfigure_console_utf8() -> None:
-    """The Windows console defaults to cp932/cp936, so writing non-ASCII / emoji raises
+    """The Windows console defaults to a legacy ANSI codepage, so writing non-ASCII / emoji raises
     UnicodeEncodeError. Forces stdout/stderr to UTF-8 + replace-unencodable-chars, so
     progress output never throws.
 

tests/test_studio_bat_ascii.py

@@ -26,7 +26,7 @@ def test_studio_bat_is_pure_ascii() -> None:
             f"studio.bat contains {len(bad_offsets)} non-ASCII byte(s) (first 5 detailed below):\n"
             + "\n".join(snippets)
             + "\n\nstudio.bat must stay pure ASCII -- cmd.exe parses it with the system ANSI "
-            "codepage before chcp 65001 runs, and UTF-8 Chinese bytes get split into garbled "
+            "codepage before chcp 65001 runs, and multi-byte UTF-8 characters get split into garbled "
             "commands. Put any non-ASCII messages inside echo, routed through a Python "
             "process that has already set PYTHONUTF8 / chcp 65001."
         )

tools/download_models.py

@@ -28,7 +28,7 @@
 import sys
 from pathlib import Path
 
-# Windows consoles using cp936/cp932 raise UnicodeEncodeError on non-ASCII / emoji; force UTF-8.
+# Windows consoles using a legacy ANSI codepage raise UnicodeEncodeError on non-ASCII / emoji; force UTF-8.
 for _stream in (sys.stdout, sys.stderr):
     try:
         _stream.reconfigure(encoding="utf-8", errors="replace")

tools/launcher.py

@@ -62,8 +62,8 @@
 # ---------------------------------------------------------------------------
 #
 # **Every string this file prints must be pure ASCII**, the same discipline as studio.bat.
-# The Windows console decodes using the system ANSI codepage (Russian cp866, Chinese cp936,
-# Japanese cp932); a frozen exe writing a single `->`-style arrow glyph is an instant
+# The Windows console decodes using the system ANSI codepage (e.g. Russian cp866,
+# Western cp1252); a frozen exe writing a single `->`-style arrow glyph is an instant
 # UnicodeEncodeError crash -- and the crash usually happens **in the error path itself**
 # (die()'s hint line), so the user doesn't see "Python isn't installed" but a PyInstaller
 # traceback instead. Comments and docstrings aren't bound by this rule (they're never

tools/mask_check.py

@@ -33,7 +33,7 @@
 import sys
 from pathlib import Path
 
-# CJK output crashes under Windows console cp932; force utf-8
+# Non-ASCII output crashes under Windows console cp932; force utf-8
 if hasattr(sys.stdout, "reconfigure"):
     sys.stdout.reconfigure(encoding="utf-8", errors="replace")
 

tools/mem_probe.py

@@ -171,7 +171,7 @@ def read(self, target_pid: int) -> tuple[float, float, float]:
 
 # ---------------------------------------------------------------- main loop
 def main() -> None:
-    # The console may be cp932/gbk or some other non-utf-8 codepage (this machine uses cp932);
+    # The console may be cp932 or some other non-utf-8 codepage (this machine uses cp932);
     # non-ASCII print output would raise UnicodeEncodeError and crash the probe itself.
     # Force utf-8 + errors=replace uniformly.
     for _s in (sys.stdout, sys.stderr):