Replace `agentcore dev --invoke "Hello"` with `agentcore dev "Hello"`.
The dev command now accepts an optional positional [prompt] argument
that invokes a running dev server directly. The --invoke/-i flag is
removed entirely — this is a breaking change for scripts using it.

Rejected: keep --invoke as deprecated alias | adds UX confusion with two ways to do the same thing
Confidence: high
Scope-risk: narrow

When running `agentcore dev "Hello"`, the CLI now automatically starts
a dev server if none is running, invokes it, and shuts it down after.
If a server is already running (via --logs or TUI), it reuses it.

- Export waitForServerReady from dev operations barrel
- Probe target port before invoking; auto-start if nothing listening
- Use Promise.race to bail early if server crashes during startup
- Silent startup (no "Starting dev server..." output)
- Helpful error when outside project with no server running
- Update tests, docs, and command descriptions

Constraint: Must not add latency when server is already running
Rejected: Always require separate terminal for dev server | poor UX parity with agentcore invoke
Confidence: high
Scope-risk: narrow

Fixes found during bugbash of auto-start feature:

- Register process.on('exit') handler to kill auto-started server,
  ensuring cleanup even when invoke helpers call process.exit(1)
- Initialize resolveServerExit with safe default to prevent
  potential undefined call if Promise executor changes
- Add unref() to SIGKILL timer in DevServer.kill() so Node.js
  exits promptly instead of hanging 2 seconds
- Use !== undefined check for prompt so empty string "" enters
  invoke path instead of falling through to interactive TUI

Constraint: process.on('exit') handlers are synchronous; cannot await server shutdown
Rejected: Modify invoke helpers to throw instead of process.exit | too many callers to change
Confidence: high
Scope-risk: narrow
Not-tested: rapid consecutive invocations (port TIME_WAIT race)

After the TCP port probe finds something listening, do a lightweight
HTTP GET to the endpoint. If the response is HTML (text/html), the
server is not an agentcore dev server — show a clear error instead
of sending a POST and dumping raw HTML as the error message.

Constraint: Cannot add a /health endpoint without modifying all Python agent templates
Rejected: HEAD request probe | some servers return different headers for HEAD vs GET
Confidence: high
Scope-risk: narrow

- Spawn child processes with detached:true and kill the entire process
  group (-pid) to clean up uvicorn workers and other grandchild
  processes on shutdown
- Fix SIGKILL fallback that never fired: child.killed is set when the
  signal is sent, not when the process exits. Use child.exitCode to
  check actual termination.
- Add explicit SIGINT handler in auto-start invoke path so Ctrl+C
  kills the server before Node exits
- Improve port identity probe: detect non-HTTP processes (e.g., raw
  TCP listeners) by treating probe failures as port conflicts instead
  of silently falling through
- Guard kill() in finally block with try-catch and remove exit listener
  to prevent leaks

Constraint: process.kill(-pid) requires detached:true so child is its own process group leader
Rejected: child.killed for exit check | set on signal send, not on actual termination
Rejected: process.exit(130) in SIGINT handler | kills Node before child cleanup completes
Confidence: high
Scope-risk: moderate
Not-tested: SIGINT during waitForServerReady phase (before invoke starts)

Reverts the detached:true, process group kill, and exitCode changes
from dev-server.ts. These affected all dev server modes (TUI, --logs,
auto-start) but were only needed for an edge case in auto-start SIGINT
cleanup. Too much blast radius for the benefit.

The command.tsx fixes (port probe, SIGINT handler, exit listener
cleanup) remain — they are scoped to the invoke path only.

…ssages

Remove auto-start logic from the invoke path. Users must now start the
dev server in a separate terminal with `agentcore dev --logs` before
invoking with `agentcore dev "prompt"`.

Improve connection error detection by adding isConnectionRefused() helper
that checks ConnectionError name, cause chain, and common fetch failure
messages instead of only checking err.message for ECONNREFUSED.

Constraint: Node fetch wraps ECONNREFUSED in TypeError cause chain
Rejected: Auto-start server on invoke | orphaned processes, complexity
Confidence: high
Scope-risk: narrow