feat: load category experiences on demand through AgentScope

[kyy-logs]

Change

Replace automatic keyword/error-triggered experience injection with agent-driven category browsing. Each framework/model/workflow directory stores JSON records and a derived compact id/name/description catalog; the agent lists a catalog and reads an applicable displayed ID using native AgentScope tools.

Preserve task-wide quotas, frozen snapshots, provenance validation and advisory-only boundaries. Add seven-field learning proposals, legacy JSON compatibility and atomic migration, trace events, and frozen multi-file evaluation. Include the previously validated grounded-contract and review-context refusal diagnostics.

Reject root or ancestor junction redirection before reads/writes and lock creation, and support long Windows migration staging/publication paths. Keep old explicit ExperienceView compatibility; the product uses the new catalog path.

Validation

• Final Windows offline pytest suite: 876 passed, 6 skipped, exit 0 (701.90s).
• Trace viewer: 8 passed, 0 failed.
• Independent non-Astra review: two Important findings repaired with RED-to-GREEN tests; repair regressions 61 passed, 2 skipped.
• No live provider calls or efficacy A/B. Direct tool calls with an extremely small 10-byte response ceiling can still return a larger error JSON; successful content responses are fully bounded.

See docs/reviews/2026-10-12-experience-catalog-acceptance.md for evidence and limits.