Problem

The retry tests in PySQLRetryTestsMixin (tests/e2e/common/retry_test_mixins.py) patch urllib3 globally (HTTPSConnectionPool._get_conn / _validate_conn) and assert the connection's own request was retried exactly N times (e.g. assert mock.call_count == 6). Under merge-queue load these flake with AssertionError because the same connection makes side-channel HTTP calls through the same mocked pool, inflating the count.

Seen on run 26867475731 / job 79234339492 (test_oserror_retries[extra_params0/1], test_retry_max_count_not_exceeded[extra_params0]) — the re-run passed, confirming flakiness. CI logs show interleaved /telemetry-ext POSTs and feature-flag GETs during the test window.

There are two leak sources, both triggered by the default enable_telemetry=True:

1. Synchronous feature-flag GET during Connection.__init__ (TelemetryHelper.is_telemetry_enabled → FeatureFlagsContext._refresh_flags → _http_client.request(GET, ...)) — deterministic, and not addressed by the prior mitigation.
2. Async telemetry POSTs to /telemetry-ext from the background executor + periodic flush thread.

The prior fix (_isolated_from_telemetry, #812) only attacked source #2, and imperfectly — so these kept flaking.

Fix (test-only, no production changes)

• Add enable_telemetry=False to the shared _retry_policy dict. It short-circuits is_telemetry_enabled before the feature-flag fetch (kills [Cleanup] Set up poetry, lockfile, mypy, and black #1) and installs a NoopTelemetryClient (kills [Cleanup] Move connector files to root #2), removing both sources at their origin without changing the retry behavior under test. Every test in the mixin merges _retry_policy into its connection params, so this single line covers all ~13 tests.
• Harden _isolated_from_telemetry() with a backstop that no-ops the actual HTTP boundary (_send_with_unified_client), closing the already-submitted-future gap for any other caller of the helper.
• Add a version_string forward-compat shim to the SimpleHttpResponse mock so the suite also runs on urllib3 ≥ 2.3 (harmless on the 2.2.x pinned in CI; surfaced during local verification).

Verification

Run against dogfood:

• test_oserror_retries and test_retry_max_count_not_exceeded — 3/3 deterministic passes for both Thrift and SEA params.
• Full PySQLRetryTestsMixin — 36 tests, all pass.

The retry behavior under test (the connection's own retried-request count) is unchanged; only the telemetry/feature-flag side-channel traffic is removed.

This pull request and its description were written by Isaac.