a plain GET on the exec path is the session list, not a failed upgrade

[lex00]

Fixes the 426 in #18.

The exec path serves two protocols

sprites-ex's Session.list_by_name/2 is literally:

Req.get(client.req, url: "/v1/sprites/#{URI.encode(name)}/exec")
# ... (Map.get(body, "sessions") || [])

spritzer registered that path as a WebSocket handler only, so an unupgraded GET fell through to websocket.Accept and came back 426 with a protocol-violation body:

level=WARN msg="exec websocket accept failed" id=fountain-0f8c1cbe-334db529
  err="failed to accept WebSocket connection: WebSocket protocol violation:
       Connection header \"\" does not contain Upgrade"

fountain asks for the session list before deciding whether to reattach or start fresh, so the 426 aborted that decision and orphaned the turn before anything ran.

An idle sprite has no sessions, so the honest answer is {"sessions": []} — which is also the answer that lets a client get on with it. A missing sprite is still a 404: "no sessions" and "no sprite" are different answers, and a client reattaching should be able to tell them apart.

The upgrade check reads both headers as the lists they are. Connection: keep-alive, Upgrade is what proxies routinely send, and an equality check would route it to the session list and break exec behind any proxy — that case is a table test.

It fixes the 426 and does not make a fountain turn complete

Verified against a real fountain deployment on k3d, with this build swapped in. The failure moves and does not go away:

before:  reattach interrupted  {"reason":"list_sessions_failed"}   # the 426
after:   reattach interrupted  {"reason":"no_active_session"}

and the actual root cause is now legible in fountain's log:

[error] GenServer {Fountain.ConversationRegistry, "..."} terminating
** (stop) exited in: GenServer.call(#PID<...>, {:write_stdin, "say hello"}, 5000)
    ** (EXIT) normal
    (fountain 0.3.0) .../conversation_server.ex:1228 ConversationServer.kick_turn/4

fountain opens the exec socket, then writes the prompt to it as stdin. spritzer's interpreter runs the command and exits immediately, so the process is already gone by the time the write lands and the conversation server crashes.

That is a separate and larger question than this PR — whether exec should hold a session open until stdin EOF rather than being one-shot. I have written it up on #18 rather than deciding it here, because it changes exec semantics for every client and is the kind of thing that should be a deliberate choice about how far the emulator goes.

go test ./... and go vet ./... green. The health endpoint now advertises both halves of the path, and a test asserts it does.