fix: retrieval_feedback table stores bloated traversal data (300-600KB per row)

[CalebisGross]

Problem

The retrieval_feedback table stores full graph traversal snapshots in the traversed_assocs JSON column. Each record is 300-600KB, making 176 rows consume ~22MB of the 116MB database (19% of total DB size for 0.5% of rows).

Current behavior

When a retrieval query runs, WriteRetrievalFeedback stores:

• traversed_assocs: Full list of every association edge visited during spread activation (300-600KB JSON)
• access_snapshot: Memory access state at query time (up to 800B)
• retrieved_memory_ids: IDs of returned memories (~400B)

This data is read back by HandleFeedback in feedback.go to apply Hebbian learning -- strengthening/weakening the specific association edges based on the quality rating.

Why it's a problem

• 176 feedback records = ~22MB. At scale this will dominate DB size.
• The traversal data is only needed until feedback is applied. After HandleFeedback adjusts the association strengths, the raw traversal is never read again.
• Most records (128/176) have empty feedback -- they stored the traversal but were never rated, so the data was written for nothing.

Proposed fix

Two options (not mutually exclusive):

Option A: Prune after feedback is applied

• After HandleFeedback processes a record, null out traversed_assocs and access_snapshot
• Keeps the query text and rating for analytics but drops the bulk

Option B: Only store traversals for rated queries

• Don't write the feedback record during retrieval
• Instead, hold traversal data in memory (keyed by query_id) with a TTL (e.g., 30 min)
• Only persist to DB when HandleFeedback is called and needs to apply adjustments
• Most queries are never rated, so most traversals never hit disk

Option C: Store only edge IDs, not full objects

• traversed_assocs currently stores full TraversedAssoc structs
• Could store just (source_id, target_id) pairs -- that's all HandleFeedback needs
• Would reduce each record from 300-600KB to a few KB

Data from pre-nuke DB (2026-03-21)

Total feedback records: 176
Rated (helpful/partial/irrelevant): 48
Unrated (empty): 128
Min traversed_assocs size: unknown (need to re-measure)
Max traversed_assocs size: ~634KB
Total retrieval_feedback size: ~22MB of 116MB DB

Backup at ~/.mnemonic/memory.db.backup-2026-03-21