fix live-status parsing and stabilize kimi harness output

[LIU9293]

Summary

• fix todo extraction in live status parsing by handling todo_write tool input across anthropic-style stream events (including input_json_delta) and normalizing todo statuses
• broaden todo.updated parsing to accept todos, items, and tasks payload aliases so wrapped stream data reliably renders the Tasks section
• harden Kimi response parsing to avoid hard-failing on empty/irregular output and fall back to usable text (with regression tests)
• regenerate live-status harness reports after parser updates and provider retries
• bump project version to 0.0.86

Testing

• bun test packages/agents/test/kimi-client.test.ts packages/agents/test/kimi-stream-status.test.ts packages/agents/test/session-inspector.test.ts packages/agents/test/qwen-stream-status.test.ts packages/live-status-harness/test/render-status.test.ts
• bun run packages/live-status-harness/scripts/generate-report.ts --layout both
• bun run packages/live-status-harness/scripts/generate-report.ts --providers claudecode --layout split
• bun run packages/live-status-harness/scripts/generate-report.ts --providers kimi --layout split

Notes

• claudecode capture now succeeds in regenerated reports
• kimi no longer fails with parser-side "empty response"; current run returns upstream 402 membership error payload from provider