fix(studio): move AI generation behind short-lived job requests

README.md

@@ -142,6 +142,7 @@ to APIMart when you request generation. The engine validates proposals and locks
 before offering them for review. Failed requests leave the draft unchanged.
 Requests allow two concurrent generations and ten starts per minute per server
 process, with two total upstream requests (including repairs and temporary-failure retries), 8192 output tokens per attempt and a 60-second total deadline.
+The Studio starts a background generation job and polls with short HTTP requests so the deployment gateway does not need to hold a connection open while APIMart responds. Cancel aborts the job. Results expire after five minutes; a server restart asks the user to generate again without changing their saved draft. Legacy API clients can still use synchronous requests.
 Without APIMART_API_KEY, generation uses local rules and reports that in the UI;
 local rules do not interpret arbitrary product descriptions. Drafts remain in
 browser storage; export before clearing browser data or moving to another origin.