fix: harden provider reliability

[N1ghthill]

Summary

• retry transient failures during provider connection validation and clarify end-to-end timing
• align Anthropic streaming tool arguments, usage accounting, tool choice, and rate-limit errors with the current Messages API
• add regression coverage, an opt-in live Anthropic smoke test, and refreshed Build Week evidence

Validation

• pnpm validate
• npm pack --dry-run --json in apps/terminuz
• 649 tests passed; 2 skipped, including the credential-gated Anthropic smoke test