fix: wait for agent_case LanceDB visibility before skill extraction

[rendigua2025-gif]

Summary

Addresses #275.

SkillClusterUpdated can arrive before the corresponding agent_case is visible through LanceDB. Previously, extract_agent_skill checked LanceDB once, raised _CaseNotYetIndexedError, and could quickly consume OME retry budget during normal cascade/indexer lag.

This PR adds a short bounded visibility wait inside _load_target_case before raising the existing retry-class error.

Scope

• Poll agent_case_repo.find_by_owner_entry(...) for a small bounded window.
• Return immediately once the target case becomes visible.
• Preserve the existing _CaseNotYetIndexedError path when the case is still not visible after the window.

This is a low-risk mitigation, not a full architectural fix. If LanceDB visibility lag exceeds the short wait window, OME retry still handles the delayed dependency. A deeper fix could emit SkillClusterUpdated only after confirmed case index visibility, carry enough case data in the event payload, or model dependency-not-ready as delayed scheduling.

Tests

• tests/unit/test_memory/test_strategies/test_extract_agent_skill.py
• tests/integration/test_ome_strategies_integration.py -k test_skill_chain_e2e

Note: I verified locally on Windows with a test-process-only fcntl stub because the current repository still imports POSIX fcntl during test collection on Windows.