Problem

The zeph_corrections Qdrant collection is consistently empty (0 points) across sessions, even after multi-turn conversations where the user provides explicit corrections or feedback.

Observed behavior

• zeph_corrections collection: 0 points (verified via /collections/zeph_corrections/points/scroll)
• user_corrections SQLite table: 0 rows
• FeedbackDetector detects correction signals, but they are not stored in Qdrant for cross-session recall

Expected behavior

When the user corrects the agent (explicit rejection, alternative request, self-correction signal), the correction should be:

1. Detected by FeedbackDetector
2. Stored in SQLite user_corrections table
3. Embedded and upserted into zeph_corrections Qdrant collection
4. Recalled in future sessions via fetch_corrections() in context/assembly.rs

Root cause hypothesis

Either:

• The write path from FeedbackDetector → SemanticMemory::save_correction() → Qdrant is not connected in the agent loop
• Or save_correction() exists but is never called after a correction signal is detected

Impact

Self-learning never improves cross-session behavior. The agent repeats the same mistakes across sessions. zeph_corrections is dead storage.

Steps to reproduce

1. Start a session, ask the agent something, get a wrong answer
2. Correct the agent explicitly (e.g. "No, that's wrong, it should be X")
3. Check zeph_corrections collection point count — still 0
4. Restart session, ask the same question — no recall of previous correction