[Feedback] Represent independent user preference state and make vector updates reversible

[theg1239]

Goal

Represent independent user preference state, apply it exactly once, and reverse it safely.

Current implementation

PR #23 now:

• stores separate interaction rows for like, save, and dislike;
• makes repeated persistent events idempotent in the normal path;
• scopes unlike, unsave, and undislike to the matching interaction type;
• filters consumed and explicitly disliked repositories from recommendation results; and
• adds transition and ranking tests.

Remaining problems

The shared database migration has the wrong primary key

ML expects (user_id, repo_id, interaction_type), but the application migration still creates (user_id, repo_id). ML currently attempts to replace the constraint at runtime, which requires DDL permission and leaves the two repositories with different schema definitions.

Relevant code:

• https://github.com/ACM-VIT/gh-social/blob/b34d89d50afe6cc222350494078ec6d5b52dec13/database/migrations/0007_add_user_feedback.sql#L1-L10
• https://github.com/ACM-VIT/gh-social/blob/b34d89d50afe6cc222350494078ec6d5b52dec13/backend/db/schema.ts#L94-L107
• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/feedback/storage.py#L45-L85

Persistence and vector application are not failure-safe

The handler commits the feedback state before updating Qdrant. If Qdrant fails, the consumer ignores the returned False, marks the event processed, and acknowledges it. A retry would not repair the vector because the database state is already idempotently applied.

• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/feedback/event_handlers.py#L154-L202
• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/feedback/consumer.py#L165-L200

Negative alpha is not an exact undo

The vector is normalized after every shift. Applying the opposite alpha to the normalized result does not restore the original vector. The current test checks the requested alpha values but does not check the resulting vector.

• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/feedback/event_handlers.py#L56-L69
• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/tests/test_feedback.py#L411-L426

The unrelated corpus target change remains

PR #23 still changes the acquisition target from 3500 to 1000.

• https://github.com/ACM-VIT/gh-social-ml/blob/819a7fe4245979065b67fdf00406ff951ea0ce3f/main.py#L41-L44

Acceptance criteria

• Like and save can coexist using the shared deployed migration.
• Unsave preserves like and unlike preserves save in the handler state model.
• Replaying an identical persistent event does not shift the vector twice in the normal path.
• Failed vector application remains retryable and is not acknowledged as processed.
• Undo restores the prior vector within a documented numerical tolerance.
• Consumed repositories are not promoted solely because they were liked or saved.
• Transition tests cover UI emission, database state, vector state, retries, and every action ordering.
• The unrelated corpus target change is removed from the PR.

Related work

• ML PR: feat: persist explicit feedback signals for ranking adjustments #23
• App feedback semantics: [Feed feedback] Treat passive swipes as neutral and emit explicit engagement events Weave#53
• App PR: feat: strengthen feed feedback handling across backend and app Weave#51
• Durable event delivery: [Feedback reliability] Persist an idempotent event queue across app and service failures Weave#56

[theg1239]

Status update after PR #23

PR #23 was merged on July 13 with additional fixes that address the acceptance criteria:

• Like, save, and dislike use independent rows.
• The shared migration now uses (user_id, repo_id, interaction_type).
• Runtime feedback-table DDL was removed from ML.
• Duplicate persistent events do not apply the vector shift twice in the normal path.
• Undo uses an unnormalized preference accumulator before normalization.
• Failed handler results are not acknowledged as successful.
• Redis replay markers and retry handling were added.
• Consumed repositories remain filtered from the ranked result.
• The unrelated corpus target was restored to 3500.
• Transition, retry, rollback, and vector restoration tests were added.

There are no GitHub Actions checks on the merged head. Recommendation: run one cross-service integration smoke test against PostgreSQL, Redis, and Qdrant, then close this issue if it passes.