fix(smoke): await the server's exit before removing its catalog, and test it

Addresses both Copilot review comments on #1978.

1. `stop()` removed the catalog dir immediately after a bare `child.kill()`,
   which only *delivers* SIGTERM. That is precisely the #1801 race
   `child-cleanup.mjs:11-17` documents, and it says both halves are needed:
   `stopChild()` closes the race on the normal path, `removeSafe()` makes the
   residual case harmless. I had used only the second. `smoke-web-app.mjs`
   already used `stopChild` for its MCP test-server child, so the helper was
   inconsistent with its own sibling. `stop()` is now async — it awaits
   `stopChild` and then removes — and all three callers await it
   (`smoke-web.mjs`'s `fail()` became async to match, so every call site had to
   `await fail(...)` or execution would run past it instead of exiting).

2. The isolation contract had no test. The smokes exit the process right after
   teardown, so a regression reintroducing the shared catalog — or dropping
   cleanup — would leave them all green. Split the two pure pieces out
   (`createTempCatalog`, `buildWebServerEnv`) and cover them in a sibling
   `prod-web-server.test.mjs`: a unique dir per run, no pre-created catalog
   file, the dir is removable, and an inherited `MCP_CATALOG_PATH` (or
   `MCP_AUTO_OPEN_ENABLED`) cannot beat the explicit one — the spread-then-assign
   ordering that guard depends on. `startProdWebServer` itself stays untested by
   design; it spawns a real launcher, which is the smokes' job.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HDdo1rNRVQnVRRdsqSrvG3
Signed-off-by: cliffhall <cliff@futurescale.com>

scripts/smoke-web-app.mjs

@@ -120,7 +120,7 @@ async function shutdown() {
     }
     browser = null;
   }
-  server.stop();
+  await server.stop();
   if (mcpServer) {
     const child = mcpServer;
     mcpServer = null;