[0x5t4l1n]

Problem

Calling HashClient.remove_server() on a configured, healthy server raises KeyError: the implementation unconditionally pops the server from _failed_clients, even when there has never been a connection failure. Consequently, callers cannot take a healthy node out of rotation using this method.

For example:

client = HashClient([("127.0.0.1", 11211), ("127.0.0.1", 11212)])
client.remove_server(("127.0.0.1", 11211))  # currently raises KeyError

Change

Make removal of the optional failure record tolerate an absent entry. Remove the node from the hasher before updating failure/dead bookkeeping, so an unknown node cannot accidentally be scheduled for reactivation after an unsuccessful removal.

Retain the existing dead-timeout recovery policy. Tests verify routing to the remaining server and reactivation after the timeout.

Validation

• Reproduced the original KeyError for tuple addresses, separate host/port arguments, and Unix socket paths.
• Add six regression cases for those three address forms with pooling enabled and disabled, plus an unknown-node bookkeeping test.
• 501 unit tests pass on Linux / Python 3.13.12.
• Flake8, Black (--target-version py39), mypy for both changed files, and git diff --check pass.

Local testing used the existing isolated environment with zstd==1.5.7.2; live-server integration and benchmark tests were deselected by the repository's unit-test configuration. Black's target override is needed because the repository lists py314 while pinning Black 24.4.2, which does not recognize that target. The six existing full-package mypy errors in the untouched AWS client module were previously confirmed on clean master; no dependency or CI configuration changes are included here.

This branch starts directly from master and is independent of #669.