feat(eval): add --serve to auto-launch the EvalServer

[kunalkushwaha]

What

Removes the two-terminal dance from agk eval. Today you must start your project in EvalServer mode in one terminal, then run agk eval in another. With --serve, AGK builds & launches the project (setting AGK_EVAL_MODE=true), waits for it to become healthy, runs the tests, and tears it down — all in one command.

This is feature B2 (auto-serve) from the FEATURES.md roadmap — the biggest eval-DX win.

Before / after

# before
AGK_EVAL_MODE=true ./myworkflow      # terminal 1
agk eval tests.yaml                  # terminal 2

# after
agk eval tests.yaml --serve          # one command

How it works

• startEvalServer runs go run . (or a custom --serve-cmd) in its own process group, so go run's compiled child is reliably killed on teardown (SIGTERM → SIGKILL after a grace period).
• waitForHealthy polls the test file's target.url /health until ready or --serve-timeout.
• Server stdout/stderr is captured and printed if startup fails (and streamed with a [server] prefix under --verbose).
• Lifecycle is signal-safe (Ctrl+C stops the server) and torn down before the os.Exit on test failure.

Flags

Flag	Default	Description
--serve	off	Launch the project in EvalServer mode for the run, then stop it
--serve-dir	.	Project directory to launch
--serve-cmd	go run .	Custom launch command (e.g. a prebuilt binary)
--serve-timeout	90	Seconds to wait for health

Testing

• go build, go vet, go test ./..., gofmt all green.
• Unit tests: parseServeCmd, and waitForHealthy against httptest (becomes-healthy and timeout paths).
• Verified end-to-end against a stub EvalServer: agk eval --serve launched it, waited for health, ran a contains test to a pass, and tore down cleanly — confirmed no lingering process afterward (process-group kill works).
• docs/EVAL.md "Run Tests" now leads with the one-command flow.

Independent branch off main. Only touches cmd/eval.go, new cmd/eval_serve*.go, and docs/EVAL.md — no README conflict.

🤖 Generated with Claude Code