Root cause of the order/seed-dependent integration failures: when the WS
upgrade's 101 response and the server's first frame (HELLO) arrive in one TCP
segment, Mint returns them as [status, headers, {:data, hello}, {:done}] - the
{:data} ordered BEFORE the {:done} that builds the websocket. process_responses
only decoded {:data} once the websocket existed, so the coalesced HELLO was
silently dropped and the client wedged in :hello_wait. Coalescing is far more
likely under full-suite scheduler load, which also starved the cadence-sensitive
tests - hence the flaky, seed-varying failures. Buffer pre-upgrade bytes and
decode them the instant the websocket is built.
Test hermeticity:
- Add Dexcord.EnvSandbox: snapshot all :dexcord app env in setup, restore
wholesale on_exit. Replaces every file's hand-rolled per-key save/restore so no
knob (timing bounds, budgets, ratelimit clock, base URLs) can leak between tests.
- gateway_fatal_test: await the directly-started supervisor's full termination
(monitor + DOWN) instead of a fire-and-forget Process.exit, so its globally-named
singletons and named ETS are gone before the next test's tree starts.
