Spike agent-side diary consolidation workflow and evals

[legreffier]

Agent-Side Consolidation Skill Spike

Summary

legreffier-consolidate is currently stronger in documentation than in
implementation. The product schema supports a useful relation graph, but the
existing consolidation story is still too server-centric and too weakly
auditable for trustworthy diary maintenance.

This spike proposes an intentional, agent-driven consolidation workflow.

Problem

• server-side clustering is weak at the distinctions that matter most:
  contradiction vs elaboration, causal chain vs temporal adjacency, replacement
  vs disagreement, implementation reference vs same-topic similarity
• current relation proposal logic is effectively “cluster => supports edges”
• relation proposals need provenance and review semantics
• dream-like maintenance is attractive, but unsafe if it rewrites entries or
  auto-accepts relations

Proposal

Treat consolidation as editorial graph curation:

• build a bounded working set instead of processing the whole diary
• generate candidate pairs intentionally
• judge each pair relation-specifically
• create proposed relations by default
• attach rationale, confidence, evidence refs, reviewer, and working-set
  metadata to every proposal
• verify value against compile/search behavior for the same task prompt

Relation policy

The skill should support:

• supports
• elaborates
• contradicts
• caused_by
• references
• supersedes

No auto-accept by default.

Dream pass

A dream pass should be a bounded background proposal loop:

• recent 20-40 entries or one subsystem slice
• one pass only
• proposal-first, never acceptance-first
• no source-entry rewriting
• no contradiction flattening

Deliverables in this spike

• rewrite .agents/skills/legreffier-consolidate/SKILL.md
• rewrite the consolidation methodology reference
• package the skill as a Tessl tile
• add eval scenarios covering bounded scope, typed proposals, contradiction
  handling, dream-pass guardrails, and verification

Follow-up work

• align the server-side diaries_consolidate endpoint with an agent-side review
  role instead of pretending to perform semantic consolidation
• define a stable metadata schema for relation proposals
• decide whether any narrow auto-accept rules are ever acceptable
• evaluate whether accepted/proposed relations should influence pack ranking
  differently