feat(typing): adaptive key geometry - learns where you tap (opt-in)

[SHAWNERZZ]

What this adds

Opt-in (off by default) adaptive key geometry: the keyboard learns where you actually land on each key and shifts its invisible hit-targets to match, so consistently off-center taps still register as the key you meant. Two independently-toggleable parts that share one strength slider:

• Learned behavior — your personal per-key landing pattern, built from both taps and swipe endpoints.
• Next-character prediction — leans tap recognition toward the letters likely to come next given the current word / suggestions.

A stats screen shows a heat-map of your learned model; an optional debug overlay visualizes the targets on the live keyboard.

Safety / non-invasiveness

• Off by default, content-free (only aggregate touch geometry — no characters or words are stored), incognito-gated, kept on-device in the existing leantype.db.
• The learned shift is capped (≤25% of the key, measured on the hit-box basis) so a clearly-on-key press can never flip to a neighbor; it ramps in with sample count; and a neighbor can only win a genuinely near-boundary tap (it must contain the touch, or be no farther than the pressed key).
• Reset is reliable — "forget my model" serializes with pending writes, so a queued sample can't resurrect deleted data.
• Lookups are O(1) from an in-memory cache; no DB work on the input hot path.

Recency window

A selectable time-decay "forget window" (1–12 months, default 3): learned geometry fades by a wall-clock half-life, so the model tracks recent typing rather than an unbounded all-time average. No schema change (reuses the existing updatedAt). The decay and cap math are unit-tested.

Notes

• Settings live under a grouped Adaptive typing section, shown only when enabled.
• Targets dev per CONTRIBUTING. Compiles standalone on dev; TouchModelManagerTest passes.