Latency-injecting gRPC test harness for race-prone flows (steer, handoff, queue drain)

Problem

Concurrency/race-heavy surfaces (steer-while-running #512 / PR #570, the active-run handoff, promote-grace, supersede/append, queue drain) are tested over an in-process bufconn with ~zero gRPC latency. The interesting interleavings (ack vs echo reordering, drain-vs-append races, terminal-race promote) all depend on timing the local harness never produces. Real deployments (TUI → remote mecated, k8s) have RTTs that make these races common.

Ask

A latency-injecting test harness for the gRPC transport so race-prone flows can be exercised under realistic delay:

• A wrapper (listener or stream interceptor) that adds configurable latency / jitter to Converse stream frames, or
• A test-only dial option to slow the bufconn path, so existing tests can be re-run under latency (a -race-style latency matrix), or
• Targeted fault-injection points (delay acks vs echoes independently) for the steer ack/echo correlation and the active-run handoff.

Goal: reproduce the class of bugs that only appear when client and server are separated by a real network (e.g. the append-vs-drain ordering, ack-after-close, promote-grace timing) deterministically in CI.

Start by locating the existing in-process bufconn tests for steer, active-run handoff, promote-grace, supersede/append, and queue drain. Compare latency or jitter injection options, then verify that the chosen harness reproduces ordering races deterministically in CI, including ack-versus-echo, append-versus-drain, ack-after-close, and promote-grace timing.