Build governed cross-run memory, retrieval, and solution reuse

[jeremymanning]

Origin

Deferred from the MVP proposed as a fresh realization of #485.

Goal

Turn the MVP's run-local journal, artifact store, summary DAG, and FTS retrieval into a durable cross-run knowledge system.

Scope

• Cross-run/project/user scopes for insights and reusable problem solutions.
• Positive and negative solution records; typed I/O compatibility plus semantic matching.
• Hybrid retrieval and reranking only after benchmarks show the SQLite FTS baseline is insufficient.
• Contradiction detection, confidence, expiry/decay, provenance, retraction, and impact tracing.
• Incremental sealing/summarization without quadratic recomputation.
• Retrieval and summary-routing evaluation on seeded-needle and task-grounded benchmarks.
• Garbage collection, retention, legal deletion/redaction, encryption, tenant isolation, and secret/PII classification.

Acceptance criteria

• Cross-run reuse improves measured success/cost without increasing corrected decomposition branching above the safe threshold.
• Negative results prevent repeated failed decomposition paths.
• Every retrieved claim resolves to immutable source spans and exact provenance.
• Retraction identifies downstream artifacts influenced by the removed insight.
• Retention/deletion behavior is documented and tested, including content-addressed data.