refactor: port typing-lag CI test to the v2 socket client

[arzafran]

What this does

The typing-lag regression test that runs in CI (tests-build-and-lag job) was the last CI-gated python test still talking to the app through the old v1 socket protocol. This points it at the v2 JSON client instead, so v1's line-protocol client can eventually be deleted without breaking CI.

Summary

• tests/test_workspace_churn_up_arrow_lag.py now imports cmux/cmuxError from tests_v2/ instead of tests/cmux.py.
• list_workspaces() / list_surfaces() row shapes are identical between the v1 and v2 clients, so the test's tuple destructuring (for _index, wid, _title, _selected in ...) needed no changes.
• The v2 client doesn't ship a send_line() convenience method (it mirrors most, but not all, of v1's API), so a small local send_line() helper was added to the test file — kept out of tests_v2/cmux.py to keep this diff scoped to one file.
• RawSocketClient, used only for the tight simulate_shortcut latency-measurement loop, was ported from v1's raw plaintext framing to v2's one-JSON-object-per-line protocol ({"id", "method", "params"} request / {"id", "ok", "result"} response), calling debug.shortcut.simulate. The warm-up-then-measure loop shape (5 warm-up calls, then N timed round trips via time.perf_counter(), same optional inter-call delay) is unchanged, so latency numbers stay comparable to prior runs.
• keep_only_first_workspace()'s retry-on-"not found" loop is preserved; v2's error text (not_found: Workspace not found) still contains "not found" case-insensitively, same as v1's "ERROR: Tab not found", so the existing match logic covers both.
• .github/workflows/ci.yml needed no changes — it invokes the same file path with the same PROGRAMA_SOCKET_PATH env var, which both clients honor.

Test Plan

• python3 -c "import ast; ast.parse(open('tests/test_workspace_churn_up_arrow_lag.py').read())" — syntax OK
• Imported the module the way CI would load it and confirmed cmux resolves to tests_v2/cmux.py (not tests/cmux.py)
• This test needs a live Programa instance and this repo's policy forbids running it locally — the `tests-build-and-lag" CI job on this PR is the real proof it still passes.