Miroir IME: A Handwriting Input Method for Android with Symbiotic HTR Dataset Collection

Prior art disclosure for Miroir IME, an Android InputMethodService replacing the standard keyboard with a full-screen handwriting surface optimized for e-ink stylus tablets (Boox Note Air 5C, 284 Hz). Original contributions: (1) dual-surface architecture — MiroirIME (universal keyboard) + CaptureActivity (standalone notepad + dataset pipeline); (2) GroupManager state machine with invariant SELECTED-never-deselected-by-system; (3) blob absorption by ray casting over per-point ellipses, with expanded bounds on selected groups; (4) double-threshold long-press (selectionDelay 300ms / editDelay 800ms) separating selection from kinetic edit mode; (5) open calibration surface — all UxK parameters user-adjustable without recompilation; (6) anonymization by structural context destruction (checksum, origin→0, dt→0) — context transformed into permanent isolation, not merely deleted; (7) multi-criteria confidence score as intrinsic dataset dimension enabling native weighted training (loss function weighting by confidence); (8) three-tier impact model: parser correction / local fine-tuning / community ODbL contribution; (9) three-tier user consent model: reject-permanent / local-only / contribute with granular scope. Related disclosures: Conduit VStar binary protocol (https://doi.org/10.5281/zenodo.20964421, CC0) and UxK gesture vocabulary (https://doi.org/10.5281/zenodo.20965531, CC0). Repository: https://github.com/nctahiti/Miroir-IME (Apache 2.0). This document is CC0.