Add Video OCR: extract burned-in (hardcoded) subtitles

[niksedk]

Implements video OCR of hardcoded/burned-in subtitles (#10944), inspired by VideOCR's pipeline.

New feature: Video → OCR burned-in subtitle...

![screenshot placeholder — drag & drop a screenshot here]

How it works

1. Scan area selection — a video frame preview with a draggable/resizable crop rectangle (custom CropAreaSelector control: move, edge/corner resize, draw-new, dimmed outside area, rule-of-thumb presets for bottom third / bottom half / full frame). A slider scrubs the preview through the video. Default: bottom third.
2. Frame extraction — ffmpeg samples the scan area at a configurable rate (default 5 fps) into near-lossless JPEGs (fps + crop + optional scale down to 720 px).
3. Frame de-duplication before OCR — frames are compared on small brightness-thresholded masks using Jaccard overlap, which follows the (bright) subtitle text rather than the moving video behind it. Consecutive near-identical frames form one group; only the middle frame of each group is OCR'ed, and frames with no bright pixels are skipped entirely as blank. This is what makes whole-movie OCR practical.
4. OCR engines
   • Paddle OCR (standalone on Windows/Linux with the existing auto-download flow, or Python) — reuses the existing PaddleOcr batch engine via a new decode-free file-input path (PaddleOcrBatchInput.SourceFileName), so the whole video is one batch: no per-image decode/re-encode, flat memory, one model load.
   • Ollama vision — e.g. the local glm-ocr model (reuses the existing OllamaOcr engine).
   • GLM API (new engine) — GLM vision models via the Z.ai / bigmodel.cn OpenAI-compatible chat completions API with API key, model and endpoint settings.
5. Line building — consecutive OCR results are merged by whitespace/case-insensitive Levenshtein similarity (default ≥ 80%) with the on-screen-longest text winning (majority vote), gaps ≤ 250 ms bridged (so a one-frame brightness flicker doesn't split a line), and blips shorter than 250 ms dropped. Vision-LLM hallucinations on textless frames (markdown fences, prompt echoes, symbol-only lines) are sanitized to empty, which correctly means "no subtitle".
6. ASSA position — optional {\anX} tag derived from the scan area position (e.g. {\an8} for a top scan area; bottom-center emits no tag).

Lines appear live in the results grid during OCR; OK loads them into the main window (with ASSA formatting when the current format is Advanced Sub Station Alpha).

Also in this PR

• PaddleOcrInstallHelper — the Paddle engine/models download prompts extracted to a shared helper (previously duplicated across view models), used by the new feature.
• Wired into the Avalonia menu, the native macOS menu bar, the shortcut system (assignable in Options → Shortcuts), DI, settings (SeVideoOcr, crop stored as percentages so it applies across resolutions), localization (LanguageVideoOcr), and docs (docs/features/video-ocr.md, F1 help).
• 32 unit tests for the merge/majority-vote/similarity/cleaning/alignment logic.

Verification

• Full solution builds with 0 warnings; all 604 UI tests pass.
• End-to-end test on a synthetic video (three subtitles burned over a moving testsrc2 background) through the real pipeline with Ollama + glm-ocr: output was exactly 3 lines with exact texts and frame-accurate timings (1–4 s, 5–8 s, 10–14 s).
• An 8-angle self-review pass was run on the diff; all 10 findings (including macOS menu wiring, GLM JSON escaping, flicker line-splitting, engine-error reporting, preview bitmap leak) were fixed and re-verified.

Closes #10944

🤖 Generated with Claude Code