SelfLearningRvfBackend feedback and contrastive learning are disconnected from retrieval

Versions

• agentdb@3.0.0-alpha.17
• @ruvector/sona@0.1.7
• Node 26.4.0, macOS arm64

Reproduction

1. Create SelfLearningRvfBackend with learning: true.
2. Insert two vectors.
3. Call searchAsync().
4. Call recordFeedback(firstHit.id, 1.0), using the returned document ID.
5. Inspect getLearningStats().

Observed:

• Returned hits expose only id, distance, similarity, and metadata; there is no query or trajectory handle.
• Feedback using the returned hit ID records zero trajectories.
• Guessing the hidden internal ID q_0 records one trajectory, proving the trajectory exists but its identifier is unreachable through the public result.
• Feeding one low-quality and one high-quality trajectory through guessed q_N identifiers still creates zero contrastive samples and zero training batches.
• forceLearn() does not change that result.

Root cause

There are three independent breaks in the loop:

1. searchAsync() creates a private q_N key but never returns it or accepts a caller-owned correlation key.
2. recordFeedback() does not assign the matching recentSearches[].quality, while contrastive sample creation requires that value.
3. ContrastiveTrainer.project() is not called by the search path, so even trained projection weights cannot affect retrieval.

Impact

The advertised self-learning path can collect neither normal asynchronous feedback safely nor use contrastive training to change rankings. Applications may believe retrieval is learning when it is effectively unchanged.

Requested fix

• Accept a caller-provided opaque feedbackId in search options, with duplicate-ID validation.
• Link feedback quality to the matching recent search.
• Apply the trained projection to retrieval only after a trained batch exists.
• Add concurrent and out-of-order feedback tests.
• Add an end-to-end held-out test showing that feedback measurably changes ranking.

Start by tracing SelfLearningRvfBackend.searchAsync(), recordFeedback(), getLearningStats(), forceLearn(), and ContrastiveTrainer.project(). Reproduce the q_N correlation and contrastive-training failures, then inspect the existing recent-search handling and retrieval path. Done means caller-owned feedback IDs work with concurrent and out-of-order feedback, contrastive batches train, and the held-out ranking test shows the trained projection affects retrieval.