Raise Peddler errors when retries are exhausted on an error status

[lineoffligbot]

Problem

With retries: > 0, exhausted retries raise HTTP::OutOfRetriesError from inside the http client, bypassing Response.wrap entirely. The error class for the same Amazon failure depends on client configuration:

Peddler.orders_v0.new(region, token).get_orders(...)             # 429 → Peddler::Errors::QuotaExceeded
Peddler.orders_v0.new(region, token, retries: 3).get_orders(...) # 429 ×4 → HTTP::OutOfRetriesError

Anyone following the README (rescue Peddler::Error, promised since v5.0: "All HTTP errors now raise Peddler::Error") silently stops catching throttles the moment they turn retries on — which is exactly when throttles matter most.

Verified against a local always-503 server: retriable alone retries correctly and raises OutOfRetriesError with the final response attached; that error then escapes Peddler's wrapping.

Fix

API#request rescues HTTP::OutOfRetriesError. When it carries the last HTTP response (OutOfRetriesError#response, populated whenever the final attempt produced a status), Peddler raises the same error that response would have raised unretried, minted by the same builder:

rescue HTTP::OutOfRetriesError => e
  raise e unless e.response

  raise Error.build(e.response)
end

Ruby's implicit exception chaining preserves the original in cause (asserted in the test):

Peddler::Errors::QuotaExceeded → cause: HTTP::OutOfRetriesError → cause: last network error, if any

Exhaustion with no response at all — every attempt failed at the network level, or the final attempt did (http.rb's performer only carries the last attempt's result) — re-raises untouched. There is no HTTP error to represent; wrapping transport failures is a separate, deliberately deferred design question.

Notes

• Technically a behavior change for anyone who discovered the leak and rescues HTTP::OutOfRetriesError around Peddler calls; filed under Fixed as it restores the documented contract, with the original error still reachable via cause.
• Raw transport errors on unretried requests (timeouts, connection resets) are unchanged and out of scope.