Replace the stub LearningEngine with real behavioral learning:

- Add SQLite table `learned_preferences` (migration 036) for cross-session
  preference persistence with UPSERT semantics
- Add `sqlite/preferences.rs` with upsert, load, and watermark-based load
  functions; enforces key ≤ 128 B / value ≤ 256 B with warn! on truncation
- Implement `LearningEngine::tick()`, `should_analyze()`, `mark_analyzed()`
  to trigger incremental analysis every 5 turns
- Implement `Agent::analyze_and_learn()`: reads new `user_corrections` rows
  since last watermark, runs regex-based preference inference, persists
  high-confidence facts (Wilson score threshold) to `learned_preferences`
- Implement `Agent::inject_learned_preferences()`: loads up to 3 highest-
  confidence preferences, sanitizes values (strip \\n/\\r), appends to the
  volatile system prompt block (after <!-- cache:volatile --> to preserve
  prompt caching, closes #1082 regression risk)
- Gate analysis on `LearningConfig::correction_detection` (independent of
  `LearningConfig::enabled` which controls skill auto-improvement)
- Preference inference uses word-boundary anchors to avoid false positives;
  covers verbosity, format, and language categories
- 40 new unit and integration tests; total test count: 6072