Recall pitfall experience on evidence-backed runtime triggers

[kyy-logs]

Previously, workflow lessons were retrieved only against Issue text and target modules, so tool/protocol lessons could be learned but never recalled. This change derives controlled kind/tool/code triggers from cited frozen evidence and injects matching advisory summaries before the next existing AgentScope model request when the corresponding event actually occurs.

The library remains frozen per task. Initial problem retrieval and later event retrieval share the three-summary/2048-byte budget, including detail reads. Existing cards retain their persisted representation and identity; injected advice stays separate from original-source evidence. Optional observation failures cannot discard tool results, and admission checks both byte capacity and the SDK context threshold without adding provider calls.

Validation:

• Full offline pytest: 827 passed, 5 skipped, exit 0 (external temporary directory).
• Trace viewer: 8 passed; wheel installation/resource checks: 3 passed.
• Fresh GPT-6.1-sol review findings reproduced and fixed with regression tests.
• Real SDK + mock HTTP lifecycle covers learning, persisted triggers, matching event and next wire-request injection.

No real-provider A/B has been run; improved reproduction success is not yet established. See docs/reviews/2026-10-12-experience-trigger-retrieval-acceptance.md for acceptance evidence and compatibility limits.