Phase 4 — Keep cleanup and empirical follow-up independent

State: open; not part of the core semantic/resolver adoption patch.

Trace-learning and source-family cleanup

Retain trace-learning as a substantive navigation facet and the source for the systems-matrix Boolean. Keep the review schema, type, template, writing skill, and tests in parity with the ### Trace-learning subsection. Correct the stale trace-derived conditional spelling when that packet runs.

Separately remove redundant x-*, github-issue, and github-pr source-family tags and stop snapshot producers from emitting them. Their platform/container classification already belongs to source metadata. Derive current inventories at execution time.

These changes are useful but do not prove tag-scope consistency. Review and commit them independently from Phases 1–3. Before workshop closure, complete them or transfer them to a named migration owner.

Agent navigation follow-up

After exact resolution and canonical heads exist, compare:

• deterministic exact membership records;
• the curated canonical head;
• query-conditioned result pointers or summaries.

Measure exact membership recovery separately from task-relevant discovery. Record wrong opens, full-artifact reads, tool calls or context cost, task outcome, and false stopping. Human search and information-foraging literature motivates this comparison but does not predict its result for LLM agents.

The trial is not a structural closure requirement unless the adopted ADR makes a retrieval-performance claim.