feat: add --prompt-retries for automatic retry on transient errors

[lupuletic]

Summary

• Add --prompt-retries <count> CLI flag to automatically retry failed prompt turns on transient ACP errors
• Add isRetryablePromptError() to classify transient vs permanent errors — retries only on ACP -32603 (internal error) and -32700 (parse error)
• Implement exponential backoff: 1s, 2s, 4s, 8s, capped at 10s
• Never retry auth, permission, timeout, or session-not-found errors
• Skip retry when the agent process crashed during the prompt
• Default: 0 retries (preserving existing behavior)

Closes #137

How it works

When a model API returns a transient error (e.g. HTTP 400/500), the ACP agent wraps it as a JSON-RPC -32603 internal error. With --prompt-retries 3, acpx will:

1. Detect the error is transient via isRetryablePromptError()
2. Wait with exponential backoff (1s → 2s → 4s, capped at 10s)
3. Re-send the prompt to the still-running agent
4. Log each retry attempt to stderr

# Enable 3 retries for flaky model APIs
acpx --prompt-retries 3 claude prompt "Hello"

# Also works with exec mode
acpx --prompt-retries 2 claude exec "Analyze this code"

Test plan

• 13 new unit tests for isRetryablePromptError() covering all error categories
• Full test suite passing (263 tests, 0 failures)
• TypeScript type check clean
• Linting + formatting clean
• Pre-commit hooks pass

Note: pnpm run check fails at the repo-wide coverage gate due to pre-existing baseline coverage totals unrelated to this change (same issue noted in PR #128).

AI-assisted PR

• Marked as AI-assisted
• Fully tested locally (pnpm build && pnpm test && pnpm lint)
• I understand what the code does

🤖 Generated with Claude Code