kapso Files

Two robustness fixes from the onboarding E2E:

• Update-crew timeout raised 240 → 360 minutes — the cap was calibrated on the default crew models; a model swap that reasons longer ran a single-trajectory lesson to within 9 minutes of it. Caps cost nothing when crews finish early. The README's Choosing models section now notes that swapped models change pacing.
• Clean teardown — the facade closes its knowledge-search clients (Weaviate, Neo4j, embeddings) at interpreter exit and whenever a backend is replaced, so scripts end without unclosed-socket ResourceWarnings; Kapso.close() is also public for callers done early.