feat: retry failed capture batches with backoff; default decision budget 5 s

[kbhatnagar1506]

What and why

• Failed capture sends were dropped on the first try, so one prod blip lost whole batches. Now a send that fails for a passing reason (connection error or timeout, 408/425/429/5xx) is retried up to 3 times with doubling backoff (0.5, 1, 2 s). The SDK honours Retry-After up to 10 s and never waits past a flush's deadline. Other statuses (401, 413, …) are not retried. A batch that still fails is dropped and counted in failed, and retries are counted in retried.
• The default decision timeout goes from 2 s to 5 s. The server allows up to 15 s per step.

Checklist

• Tests added or updated (retry until it lands, bounded with the drop counted, no retry for non-passing statuses, Retry-After cap, flush deadline)
• Commit messages follow Conventional Commits (fix:, feat:, ...)
• The SDK still never raises into the caller and still fails open

🤖 Generated with Claude Code

https://claude.ai/code/session_018ruNgP5DUwEkpSrXnv5R8T

---

Generated by Claude Code