Use model-only word prediction

[enaboapps]

Closes #880

Change

• Use the pinned on-device model as the sole word prediction engine. The saved enhancedWordPrediction field remains compatible but no longer changes engines.
• Load it in the worker asynchronously; leave suggestion slots empty while loading and show an unavailable status on failure or slow inference while keeping keyboard input usable.
• Remove the bundled lookup, SQLite source, converter, package entry, and CI checks. Keep the tokenizer and blocklist.
• Use one Word prediction control and update its help text.

Validation

• Node 24: npm run lint, npm test (224 UI tests, 5 feed tests), npm run build passed.
• Rust 1.97.1: cargo fmt --manifest-path src-tauri/Cargo.toml --check, cargo clippy --manifest-path src-tauri/Cargo.toml --all-targets -- -D warnings, cargo test --manifest-path src-tauri/Cargo.toml (552 unit tests, 7 config tests) passed.
• Model quality comparison, 6 synthetic contexts: model-only returned 5 suggestions for w, wa, cup of, ket, wh, and fa; tea, kettle, WhatsApp, and water remained present. Model-only inference took 60–353 ms; prior combined model samples took 32–156 ms and lookup-only median was 0.01 ms after a 6.7 s startup. The ket tail differs (model-only includes ketogenic, keto, kets rather than lookup words), while its top kettle result remains. More real-use validation is pending.
• Windows 11 Parallels ARM64: built src-tauri/target/aarch64-pc-windows-msvc/debug/switchify-pc.exe from this branch and launched it in the interactive desktop session. The existing Space Select assignment started scanning. A temporary F8 Open keyboard assignment opened the native scanned keyboard; its first frame showed Loading predictions · Keyboard ready and five empty suggestion positions. On resuming the scan, the status advanced to Letters · Select a row, which means model loading completed without a failure status. Escape closed the overlay. The temporary F8 assignment was removed. The VM test used injected keyboard switch events; no physical switch device was connected. It did not verify a completed model suggestion in a Windows text app.
• Independent review of bd33b340e9806478c40917bfd8516799f8448e87: no actionable findings. All required CI passed, including frontend, macOS native, Windows native, CodeQL, JavaScript/TypeScript analysis, and dependency audit.

Follow-up: prediction feedback and retry

• The keyboard header now shows a passive prediction badge alongside the scan prompt: loading, ready for input, no suggestions, suggestion count, paused tracking, or unavailable. It contains no typed text and cannot be scanned or selected.
• On failure, the toolbar adds Retry predictions. Selecting it discards the old worker and private context, starts loading a fresh worker, and keeps the keyboard open. The setting and help text explain this.
• Node 24 lint, tests (224 UI + 5 feed), and build passed. Rust 1.97.1 fmt, clippy with warnings denied, and tests (555 unit + 7 config) passed. Native frame, retry state, and private-state cleanup tests were added.
• The new badge and retry tile were verified with native-frame and fake-state tests; the earlier Windows VM interaction above was on the preceding head. No additional live VM interaction has been performed for this follow-up.
• Independent review of latest head 772ba6e: no actionable findings. All latest-head CI passed, including Windows UI Access and macOS native packaging.