Add offline Chinese->English translation support and improved inplace layout

[itrejomx]

Summary

• add Chinese -> English offline mode with RapidOCR + OPUS-MT zh-en
• add language-aware status labels and safer source-language normalization
• improve inplace overlay layout by stacking overlaps and ordering labels for vertical reading flow
• refresh README intro and language support docs

Testing

• pytest tests -q
• python3.11 -m py_compile src/interpreter/config.py src/interpreter/gui/main_window.py src/interpreter/overlay/base.py

[greptile-apps]

Greptile Summary

This PR adds Chinese→English offline translation (RapidOCR + OPUS-MT zh-en) alongside the existing Japanese pipeline, introduces a language selector in the GUI, and refactors the inplace overlay to stack overlapping labels for vertical reading flow.

• P1 — inplace overlay swaps labels: arrange_overlay_regions sorts regions by (-x, y) internally, so the returned arranged list is in a different order than label_entries. The index-based zip(label_entries, arranged) on line 411 of base.py then assigns each QLabel the position of a different region — in the two-region case the two labels are fully swapped on screen.

Confidence Score: 3/5

Not safe to merge as-is: the inplace overlay label-position bug will visibly swap translated text across regions whenever more than one caption is on screen.

One confirmed P1 (label/position index mismatch after internal sort) pulls the score below the P1 ceiling of 4. The P2 findings (CUDA leak, warnings scope) are non-blocking but worth fixing before shipping.

src/interpreter/overlay/base.py (P1 label mismatch), src/interpreter/translate.py (P2 CUDA/warnings scope)

Important Files Changed

Filename	Overview
src/interpreter/overlay/base.py	Adds arrange_overlay_regions for overlap-aware inplace label layout, but contains a P1 bug: arranged is returned in sorted (-x, y) order while label_entries retains insertion order, causing index-based zip to assign each label the wrong position.
src/interpreter/translate.py	Refactors into JapaneseTranslator/ChineseTranslator with a language-aware Translator wrapper; warnings.catch_warnings() scope exits before the heavier AutoTokenizer.from_pretrained call, and the CUDA fallback path leaves a GPU translator object allocated before overwriting it.
src/interpreter/ocr_rapid.py	New ChineseOCR backend using RapidOCR; duplicates the bgra_to_rgb helper already defined in ocr.py (previously flagged).
src/interpreter/ocr.py	Renames OCR to JapaneseOCR and adds a language-aware OCR façade that delegates to JapaneseOCR or ChineseOCR; straightforward delegation pattern with no issues.
src/interpreter/config.py	Adds SourceLanguage enum, normalize_source_language, and model-name helpers; serialization round-trip and fallback handling look correct.
src/interpreter/gui/main_window.py	Adds language combo, dynamic model-name labels, and extracts _restart_process_worker; the _fixing_ocr flag is still not set for the Chinese OCR failure path (previously flagged).
src/interpreter/gui/workers.py	Adds contains_han, set_source_language, and language-gated translation-skip logic; clean and correct.
tests/test_overlay_layout.py	New tests verify arrange_overlay_regions return values by text key but don't catch the index-based zip mismatch in the caller that causes the P1 bug.
tests/test_translate.py	Adds boilerplate version patching; existing translate tests unchanged and look correct.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Screen Capture] --> B{source_language}
    B -->|JAPANESE| C[JapaneseOCR / MeikiOCR]
    B -->|CHINESE| D[ChineseOCR / RapidOCR]
    C --> E[OCR regions list]
    D --> E
    E --> F{contains_japanese / contains_han filter}
    F -->|pass| G{Translator wrapper}
    F -->|skip| H[emit empty]
    G -->|JAPANESE| I[JapaneseTranslator / Sugoi V4]
    G -->|CHINESE| J[ChineseTranslator / OPUS-MT zh-en]
    I --> K[translated regions]
    J --> K
    K --> L{overlay_mode}
    L -->|BANNER| M[BannerOverlay]
    L -->|INPLACE| N[arrange_overlay_regions sort by -x,y]
    N --> O[zip(label_entries arranged) order mismatch]
    O --> P[InplaceOverlay QLabels]

Loading

_{Reviews (2): Last reviewed commit: "chore: quiet Chinese model startup warni..." | Re-trigger Greptile}

[greptile-apps]

_fixing_ocr never set for Chinese OCR failures

When OCR fails in Chinese mode and the user clicks "Fix Models", _fixing_ocr remains False because there is no branch for SourceLanguage.CHINESE. The worker restarts and the ocr_status signal fires "loading", but the status label path that shows "Downloading..." (guarded by _fixing_ocr) is skipped, so the label jumps directly from its error state to "Loading...". Setting self._fixing_ocr = True when Chinese OCR is in the failed set would keep the labels consistent.

[greptile-apps]

Duplicate import importlib.metadata

importlib.metadata is imported on line 5 and again on line 13. The second import is redundant. The same pattern appears in tests/test_language_backends.py (lines 5 and 13 in that file).

[greptile-apps]

Label-to-position mismatch due to internal sort in arrange_overlay_regions

arrange_overlay_regions sorts the input regions list internally by (-x, y) before computing positions, so the returned arranged list is in a different order than label_entries. The zip(label_entries, arranged) on line 411 pairs by index, meaning label for region A ends up receiving the position computed for region B.

Concrete example: two regions at x=100 and x=200 — label_entries keeps the original order [A(x=100), B(x=200)], but arranged is [B_layout, A_layout] (sorted by -x). The zip assigns label_A.move(B_layout) and label_B.move(A_layout), swapping both overlays on screen.

Fix: match by text field rather than relying on positional alignment:

layout_by_text = {item["text"]: item for item in arranged}
for label, text, _ in label_entries:
    if text in layout_by_text:
        layout = layout_by_text[text]
        label.move(layout["x"], layout["y"])
        label.show()
        self._labels.append(label)