enhancement(core): LearningEngine is a stub — no actual behavioral learning from interaction patterns

[bug-ops]

Problem

LearningEngine (crates/zeph-core/src/agent/learning_engine.rs, ~94 lines) exists as a structural placeholder but implements no actual learning logic. The agent does not improve its behavior based on interaction history.

Current state

The file contains:

• A struct with a single reflection_used flag
• No pattern analysis
• No preference inference
• No behavioral adaptation between turns or sessions
• No connection to FeedbackDetector outputs or user_corrections

What a functional LearningEngine should do

1. Preference inference: detect recurring user preferences from corrections (e.g. prefers concise answers, always wants code in Rust, dislikes emoji)
2. Tool usage patterns: learn which tools the user relies on and pre-suggest them
3. Skill affinity: track which skills produce high-quality outcomes for this user (connect to skill_outcomes table)
4. Cross-session adaptation: persist learned preferences to zeph_key_facts or a dedicated user_preferences collection
5. Response style adaptation: adjust verbosity, language, format based on feedback history

Affected tables (currently empty/unused)

• skill_outcomes — 0 rows
• skill_usage — 0 rows
• user_corrections — 0 rows (also blocked by fix(memory): zeph_corrections Qdrant collection never populated — self-learning corrections not persisted #1910)
• skill_versions — 0 rows

Impact

• Self-learning feature is advertised but non-functional
• No behavioral personalization across sessions
• FeedbackDetector signals are detected but never acted upon beyond skill re-ranking

Suggested approach

Start with the simplest useful behavior: after each session, scan user_corrections for patterns and write 1–3 preference facts to zeph_key_facts via memory_save. Build from there incrementally.