test: clean up transport recovery timers

[danyel117]

Summary

• dispose each transport-error test plugin instance so scheduled recovery timers cannot outlive afterEach
• stop scheduling retries in the classification loop by using the one-failure ceiling
• widen polling deadlines to tolerate loaded CI runners while preserving bounded assertions

Context

The publish workflow for #25 repeatedly completed the new recovery tests successfully, then failed when leaked timers wrote after their test directory had been removed.

Validation

• 20 focused stress runs passed
• bun run lint
• bun run typecheck
• bun test (190 pass)
• bun run build
• bun run pack:dry-run
• Claude Opus review: no merge blocker