adapter: add LaunchDarkly reconnect integration test

[jasonhernandez]

Motivation

incident-984 was a runtime failure: the LaunchDarkly data source stopped reconnecting after its streaming connection dropped with a non-Eof error, silently wedging flag sync. The existing test/launchdarkly nightly covers value sync, persistence, targeting, and the kill switch, but nothing exercises reconnect after a mid-stream drop. A prior attempt to add such a test was abandoned because the failure couldn't be reproduced against real LaunchDarkly ("the test would need to cut the connection"). This reproduces it deterministically against a mock.

This is the reconnect test from #37026, un-stacked from the SDK bump (#37025) and rebased onto current main, so it lands ahead of the upstream-SDK migration. The goal is to simplify review and increase confidence: the test characterizes reconnect behavior against today's (fork) SDK first, and the SDK bump must keep it green.

What changed

Production: a hidden --launchdarkly-base-uri flag (env LAUNCHDARKLY_BASE_URI) that overrides the SDK's streaming/polling/events endpoints with a single base URL, via the SDK's ServiceEndpointsBuilder::relay_proxy. Generally useful for relay-proxy setups; here it lets tests point the SDK at a mock. Threaded through SystemParameterSyncClientConfig::LaunchDarkly → ld_config. The current fork SDK (2.6.x) already exposes this API, so no SDK change is needed.

Test (test/launchdarkly-reconnect):

• mock_ld.py — a minimal mock of the LD streaming API. The first streaming client gets an initial flag value (2 GiB), then the connection is reset mid-stream with a TCP RST (SO_LINGER 0) so the SDK sees a non-Eof transport error (the incident class), not the Eof it always recovered from. Every reconnecting client gets an updated value (3 GiB).
• mzcompose.py — boots environmentd pointed at the mock (mapping max_result_size to the flag) and asserts SHOW max_result_size reaches 3GB. That can only happen if the data source reconnected after the reset; a regressed SDK stays stuck at 2GB and the assertion times out.
• Wired into the nightly pipeline. Needs no real LaunchDarkly credentials (unlike test/launchdarkly).

Validation status

• cargo check clean (mz-adapter, mz-environmentd, mz-sqllogictest); bin/fmt clean.
• The test was previously validated green against the fixed upstream 3.1.1 SDK (in the stacked form on adapter: add LaunchDarkly reconnect integration test #37026). It has not yet been run against the current fork SDK on main. It should pass (we reverted to the fork precisely because it handles incident-984's non-Eof reconnect), but the mock uses a TCP RST specifically and that path is unverified on the fork.
• Draft until the nightly LaunchDarkly reconnect job is confirmed green on this branch (fork SDK). Trigger via the ci-nightly label.

🤖 Generated with Claude Code