fix(harness): recover stalled and failed agent runs

[touch2be]

Summary

• make pending follow-up removal idempotent when dispatch wins the UI race
• abort unresolved agent-tool work promptly and return connected-tool failures to the model instead of failing the run
• keep shell timeout and force-kill timers alive so background process groups cannot stall sessions indefinitely
• classify provider quota, network, connection, and WebSocket failures as actionable retryable errors

Regression coverage

• repeated follow-up removal
• unavailable connected tool continuation
• background shell process timeout
• provider limit and transport error normalization

Test plan

• pnpm vp test packages/harness/src/harness.test.ts packages/harness/src/agent-tools.integration.test.ts packages/harness/src/models/transport-contracts.test.ts
• pnpm vp check