Retry transient 5xx and network errors with exponential backoff

[jasiek]

Summary

Upstream genealogy services (Geneteka, GenBaza, Lubgens, Genealogia w Archiwach, GenPod) intermittently return 502/503/504 or briefly drop connections. Today these surface as tool failures even though a short retry almost always recovers. This PR adds a shared transient-error retry layer used by every live HTTP client.

What changed

• New src/polish_genealogy_mcp/sources/_http_retry.py — request_with_retry() wraps an HTTP send callable with full-jitter exponential backoff. Retries on:
  • httpx.TransportError / httpx.TimeoutException
  • HTTP 502, 503, 504
    Defaults: 3 attempts, 1s base delay, 30s cap. Non-retryable responses are returned untouched (the caller still calls raise_for_status).
• All five live clients now route their requests through the helper:
  • geneteka (search, get_regions_html)
  • genbaza (_get)
  • lubgens (search)
  • genealogia_w_archiwach (Vaadin bootstrap GET/POST + _post_uidl)
  • genpod (login GET/POST, post-login home GET, GraphQL POST)
    The per-source rate limiter's wait() is invoked inside the retried callable so pacing is still respected across retries.
• Configurable via env, CLI, and Claude Desktop user_config:
  • GENEALOGY_RETRY_MAX_ATTEMPTS (default 3)
  • GENEALOGY_RETRY_BASE_DELAY (default 1)
  • GENEALOGY_RETRY_MAX_DELAY (default 30)
    Wired through _cli_config.CONFIG_ENTRIES and manifest.json.

Tests

• New tests/test_http_retry.py: success on first try, retry-then-success on 502 and ConnectError, exhaustion on persistent 503 / ReadTimeout, no retry on 404, jitter respects max_delay.
• tests/test_genbaza_client.py: existing 503 case keeps the HTTPStatusError assertion (now with GENEALOGY_RETRY_BASE_DELAY=0 to keep the test fast); added a 502 → 502 → 200 end-to-end retry case and a non-retried 404 case.
• Full suite: 119 passed.

Reviewer notes

• The helper deliberately accepts an injectable sleep so unit tests don't actually sleep; production code uses time.sleep.
• Full jitter (random.uniform(0, capped_delay)) is used to avoid synchronized retries if multiple requests fail at once.
• Each retry passes through the rate limiter — so for a chatty source with a 5s pacing window, a retry will still wait until the next slot. That is intentional: we'd rather be slow than 429'd.

🤖 Generated with Claude Code