Design: first-class relation edges between facts (motivated_by / applies_to, beyond supersedes)

[minerva-sky]

Third of the memory-model design trio (provenance #13, lifecycle #14). Design only — no code yet.

What exists

The schema is already ahead of the code. fact_links (001 migration) is a generic directional edge table with a free-form link_type, and SQLiteStore#insert_fact_link accepts any type. But only two types are ever written:

• supersedes — Resolver, on supersession (resolver.rb:276)
• conflicts_with — conflict pathway

Core::FactGraph BFS-traverses exactly those two via hardcoded discover_supersedes / discover_superseded_by / discover_conflicts methods. Nothing else in the system can create or read another edge type.

The gap

Facts carry no why and no where. Two relations would earn their keep:

1. motivated_by — preference → the fact that explains it. "Val prefers squash merges" motivated_by "linear history requirement on ups.dev". Recall of the preference without its motivation invites the agent to over-generalize it; and when the motivating fact is superseded or expires (Fact lifecycle: expiry-unless-reaffirmed + ratification surface #14), the preference should be flagged for re-review, not silently persist.
2. applies_to — fact → scope fact (a project, a repo, a person). Today scoping is a column (project scoping, 002); an edge generalizes it to non-project scopes and lets one fact scope several others.

Proposed design

• No schema change. fact_links.link_type already takes any string. Add a VALID_LINK_TYPES constant (supersedes, conflicts_with, motivated_by, applies_to) validated in insert_fact_link, so typos don't create orphan vocabularies.
• Creation paths (in priority order):
  1. MCP management tool link_facts(from_id, to_id, type) + inverse unlink — cheapest, lets the agent record relations it infers during conversation.
  2. Distill-time: the distiller prompt emits optional relations: [{to: <sibling index>, type: ...}] between facts extracted from the same content item. Same-batch only — no cross-corpus matching at distill time.
• Traversal: collapse FactGraph's per-type discover methods into one generic discover_links_of_type over VALID_LINK_TYPES; edges gain direction labels for the formatter.
• Recall surfacing: when a recalled fact has outgoing motivated_by edges, append the target fact as one-line context ("because: …"). applies_to filters rather than annotates.
• Interlock with Provenance + authority on facts; keep and surface contested claims #13/Fact lifecycle: expiry-unless-reaffirmed + ratification surface #14: edges get the same provenance columns proposed in Provenance + authority on facts; keep and surface contested claims #13 (who asserted, when); lifecycle transitions in Fact lifecycle: expiry-unless-reaffirmed + ratification surface #14 walk incoming motivated_by edges and mark dependents needs_review when a motivating fact expires or is superseded.

Non-goals

• Free-form/user-defined edge types (vocabulary stays closed until a real need shows).
• Graph queries beyond the existing BFS depth-5 (no path queries, no transitive closure).
• Backfilling relations over the existing corpus with an LLM pass — separate proposal if ever.

Implementation is gated on reaction to this trio; a tracked task exists for it on my side.