Reconsider adaptive touch correction? Slimmer proposal with details #676

Thanks again for reviewing #671. I'd like to explain the feature in more detail and offer a much slimmer version, because I think it addresses your concerns.

Why it matters
I'm an experienced, heavy phone typist. Still, over ~1,250 recorded touches about 13% landed on a neighboring key — roughly every 8th letter. And it's systematic, not random: "a" hit as "s" is consistently my most frequent miss. A fixed, per-user offset like this is exactly what adaptive touch correction fixes, and as far as I know it has been standard in mobile keyboards for many years.

What it does
The keyboard learns where this user actually touches each key, based on the words they finally commit (including reverted autocorrections and self-corrections). Once a key has enough data (30+ touches), a touch near the border between two keys is assigned to the key the user most likely meant. The shift is capped to a fraction of a key width, and the dictionary receives the corrected coordinates too.

Your concerns

• Determinism: same data, same key. It only differs from today near key borders and only after enough data. With the feature off, the exact current code path runs.
• Latency: per touch a few arithmetic operations on cached offsets, no I/O.
• Muscle memory: it adapts to the user's muscle memory instead of changing it.
• Code footprint: of the ~1.5k lines, ~1.45k are in a separate package (model, statistics screen, 17 tests). Core changes are ~115 lines.

Slimmer proposal

• One single switch ("Adapt to my typing"), fixed strength, no other settings.
• Move all recording logic out of InputLogic into the separate package, leaving ~60 lines of hooks in core.
• Optional: drop the statistics screen if you prefer.

If you're still not convinced, no problem — I'll keep it in my fork.