gap: LearningExtractor unused — observations never flow into learnings

[cmbays]

Problem

123 observations exist across 41 runs. Zero learnings have citations linking them to observations. LearningExtractor.ts exists in the codebase but is never called during cooldown.

All 7 learnings were manually recorded by the cooldown LLM synthesis, not extracted from observation patterns. The observation-to-learning pipeline is missing.

What needs designing (needs-human)

1. What patterns in observations should trigger a learning? (e.g. repeated friction of same taxonomy = learning candidate)
2. Should extraction be automatic during cooldown, or require human review before promoting to knowledge store?
3. Should citations be backfilled on existing learnings that reference the same run observations?

What needs building (once designed)

• Wire LearningExtractor into the cooldown prepare or complete path
• Populate citations[] when a learning is created or updated
• Test that a new cycle's observations produce at least one learning with a citation

Current state

LearningExtractor.ts: exists, appears unused
learning.citations[]: always empty
learning.derivedFrom[]: always empty