[Buyer Gap] Persist and explain Event Lineage channel evidence

[seonghobae]

Buyer problem

reconstruct() already computes the winning edge's per-channel scores (temporal, secondary_key, text, optional llm), but the production persistence contract deliberately collapses each edge to (parent_post_id, child_post_id, fused_score). /api/lineage and the Buyer DAG therefore expose only the fused score.

A buyer can see that two posts were linked but cannot answer:

• which independent signals supported the edge;
• whether the optional LLM channel participated;
• which signal dominated the decision;
• whether an edge was created from weak agreement across channels or one strong signal;
• how to audit a later reconstruction after model or weight changes.

This is a distinct follow-up to the #264 ontology relationship-path repair. Do not mix the two: #264 explains typed Knowledge Graph paths; this issue explains reconstructed post-to-post lineage evidence.

Verified causal boundary

• lineageweave.models.Edge retains channel_scores.
• lineageweave.reconstruct._best_parent returns the winner's channel scores.
• lineageweave.lineage_persistence documents a three-column persistence projection.
• backend.app.lineage_ingestion.persist_lineage_edges writes only fused_score.
• post_lineage_edge stores only fused_score.
• visible_lineage_graph and frontend/src/api.ts::LineageGraphEdge return only fused_score.

Required design

3NF persistence

Add an effective schema migration with a child table such as:

post_lineage_edge_signal
- parent_post_id
- child_post_id
- signal_code
- signal_score
- signal_weight
- signal_contribution

Requirements:

• composite FK to post_lineage_edge with ON DELETE CASCADE;
• one row per edge and active signal;
• controlled lookup values for temporal, secondary_key, text, and llm;
• exact decimal constraints for score/weight/contribution;
• no JSONB for billable/auditable numeric facts;
• no synthetic llm row when the LLM channel is unavailable;
• weights must be the normalized active weights actually used for that reconstruction;
• contribution must reconcile with the persisted fused score within an explicit numerical tolerance.

API and Buyer surface

Return an additive, typed channel_evidence collection on each visible lineage edge. Each item must include:

signal_code
signal_label
score
weight
contribution
rank

The Buyer surface must provide:

• an accessible edge-detail disclosure;
• exact values without hover-only interaction;
• a plain-language explanation that the relation is inferred, not causal;
• a clear indication when no LLM channel participated;
• deterministic ordering by contribution, then controlled signal order;
• print/export preservation of the same values.

Provenance and versioning

Persist or reference the reconstruction version, active weight profile, and generated-at time so a later rebuild cannot silently rewrite the meaning of historic evidence. PostgreSQL remains authoritative; PROV-O/RDF export is a projection, not the source of truth.

RED → GREEN acceptance tests

• A reconstructed edge with four channels round-trips all four scores and normalized weights through PostgreSQL and /api/lineage.
• A no-LLM reconstruction persists exactly three channels and never fabricates an LLM score.
• Signal contributions reconcile to fused_score within the documented tolerance.
• Rebuilding replaces edge and signal rows atomically; no orphan or stale signal rows remain.
• Duplicate rebuilds are idempotent.
• ABAC filtering never reveals evidence for an invisible endpoint.
• The UI renders exact values with keyboard and screen-reader access.
• The UI labels the edge as inferred and does not state causality.
• Migration clean-install and upgrade rehearsal pass.
• Production statement/branch coverage and public API docstrings remain 100%.

Non-goals

• Do not re-score or reinterpret old edges without an explicit reconstruction run.
• Do not make one channel authoritative.
• Do not expose prompts, responses, credentials, or private source text in the evidence payload.
• Do not fold this into the current ontology-path repair or the GNB navigation behavior.