Align LeanTypeDual with current upstream and retire retained backends

[AsafMah]

Purpose

Promote the installed, upstream-aligned LeanTypeDual EXP baseline into dev. Preserve the working keyboard while reducing intentionally retained older upstream implementations. This is not a release or a claim that the remaining fork is already minimal.

Scope

• Integrate LeanBitLab/LeanType through 52567c0f1e52335583e4abfeb312eb6af0d7f70f, preserving upstream ancestry.
• Keep live dual-thumb/manual/timed composition, the established connector, fragment backspace, pointer-ID normalization, app identity and stored user data.
• Remove the retained Java gesture fallback, bundled llama backend, retained host handwriting runtime, and separate Offline Lite distribution. Use upstream's native gesture/plugin paths and three-flavor on-demand dictionary packaging.
• Include the already-integrated editor ownership, runtime, import and dictionary-state corrections tracked in Correct keyboard runtime defects found during the v4.2.0 integration review #152, Fix imported two-thumb settings and gesture dictionary state staying stale #153 and Repair inherited AI request ownership and offline generation/state defects #154. Their bundled-fallback/native-only portions are superseded by backend retirement, not maintained as private engines.
• Remove dead two-thumb scaffolding and update product/release gates and documentation. Other optional fork extras remain unchanged pending categorization.

The diff against dev also contains the earlier v4.2.0 integration and corrections, not just the final alignment commits. Principal boundaries: integration f1c0fb71d, retirement b0398220a, initial user feedback b1694d8df.

Remaining fork differences

The categorized inventory is recorded in docs/HANDOFF.md. Its frozen comparison is b1694d8df against upstream 52567c0f, not this PR's diff against dev: 282 paths, with approximately 65% of added text in tests, tooling and documentation. The inventory separates dual-thumb core, retired backends, optional extras, later carried repairs and small upstream candidates. No additional optional feature was removed or upstream PR submitted during this inventory pass.

Evidence

• 35 targeted upstream-merge cases, including OCR relayout.
• 187 native-routing/composition/settings cases and 59 offline-plugin/language/editor-ownership cases, with fresh-result/baseline gates.
• 40 repository-tooling tests and the three-flavor source contract.
• Test-only CI follow-through: 7b8527bee bounds the Robolectric worker heap; 143416bb2 shares and resets the input-method fixture across application startups. Four new cases cover actual shortcut/subtype queries and automatic fixture reset on SDK 33 and 35. Two fresh full Windows runs contain 610 cases and only the four existing Parser baseline failures. No baseline/exclusion changes or production exception suppression; all diagnostic probes removed. The GitHub checks show the result for the current head.
• Standard, Standard Full and Offline EXP APKs assembled; package/SDK/permission/ABI and unbundled-asset contracts inspected.
• Standard EXP installed with data-preserving replacement; selected normal keyboard unchanged. Existing gesture library and English dictionary preserved. User's initial feedback: "feels ok". The subsequent CI and inventory follow-ups do not change this installed APK.
• Installed APK SHA-256: 8291f4fca9fd4bffea57c497d7cfedc2046c1d670245f091c1b80a404c78fe2f.

Limits

• No agent-driven phone swipe test: the user explicitly chose installation without switching keyboards. Initial feedback is not exhaustive gesture/native-plugin validation.
• The original Notes/Word paragraph-deletion report remains unresolved.
• External plugins retain their own runtime/privacy boundary. No private llama rebuild is pending.
• No version bump, release publication, automatic merge, or further optional-feature removal is included.

Related: #16, #148, #152, #153, #154. Historical merged #134 is not this integration PR.