Learning Mode — adaptive bigram training (opt-in)

[JaraLucan]

Idea

An opt-in "Learning Mode" that helps you target your slowest key combinations.
It measures inter-key latency for each letter pair (bigram), then biases word
generation toward the pairs you're slowest at, and shows a small analytics panel
after the test. Inspired by the FlowType concept, rebuilt natively for Monkeytype.

What it does (only when enabled)

• Tracks average inter-key latency per bigram, stored locally, separately per
  language + keyboard layout (latencies aren't comparable across them).
• Adaptive word generation: ~60% of words are weighted toward your slowest
  bigrams, ~40% stay random so tests stay varied. Falls back to normal random
  generation until there's enough data (~10 reliable pairs).
• Post-test "learning panel": your 5 slowest pairs this test, each with an
  improvement/decline indicator vs your all-time average, plus consistency and a
  plateau hint.
• An overview modal (graduation-cap nav button) with all-time slowest pairs,
  a WPM trend sparkline, pause/resume recording, and clear-data.
• Plateau detection (last 8 sessions within a 10 wpm range → advice).

How it differs from the existing "weak spot" funbox

weak spot is per-character, in-memory, per-session, with no persistence or UI.
Learning Mode is per-bigram, persistent, scoped per language/layout, with
adaptive generation and an analytics UI. It's additive, not a replacement.

Principles

• Opt-in, off by default — zero behavior change when disabled.
• No backend changes — all data in localStorage; the two new config booleans
  ride the existing config schema/sync.
• No gamification, minimalist UI using existing theme tokens.

Scope / limitations (v1)

• Learning data is device-local (not account-synced) by design for v1.
• Adaptive generation applies to normal word/time modes (not quote/zen/funbox).

Status

Working implementation on my fork, all frontend tests + lint + format green,
42 new unit tests. Happy to adjust scope/UX based on feedback before opening a PR.

[Screenshots + live demo link here]

Would a feature like this be welcome? Any concerns before I open a PR?