Test harness stub hard-codes port 8787, flaking under concurrent runs

[justin808]

Problem

AgentCoordTest#test_http_integration_harness_preserves_backend_test_failure_status (test/agent_coordination_cli_test.rb:1176) fails with Errno::EADDRINUSE ... 127.0.0.1 port 8787 whenever any other process holds that port. The harness's stub npx hard-codes 127.0.0.1:8787 (test/agent_coordination_cli_test.rb:73-78, introduced in PR #24) with no free-port allocation, so concurrent lanes on one machine (a second test run, a real wrangler dev, another agent worktree) make the suite flake.

Evidence (QA lane, batch ac-datafix-20260715-2339, 2026-07-17)

• Failed once during a full-suite run while another lane was active: Expected: 42 Actual: 1, "wrangler dev exited before becoming healthy", EADDRINUSE on 8787; clean rerun passed (12 suites, 426 runs, 0 failures).
• Deterministic two-way repro: hold the port with a dummy TCPServer → identical failure signature; release it → passes.

ruby -e 'require "socket"; TCPServer.new("127.0.0.1",8787); sleep 60' &
bundle exec ruby -Itest test/agent_coordination_cli_test.rb -n test_http_integration_harness_preserves_backend_test_failure_status

• Pre-existing: present and failing-when-port-busy at heads before PRs Track machine and session identity via environment variables #68/Normalize heartbeat status vocabulary at the write path #71; not a regression from the identity/status work.

Direction

Allocate an ephemeral port in the harness stub (bind port 0, read the assigned port, pass it through to the stub server and health check) so parallel suite runs and concurrent wrangler processes cannot collide.