test(neural-live): budget for a cold 95MB download

Production verified working by direct capture (vocalized output, WASM
fallback after the expected WebGPU warnings); the failure was a
fresh-profile download exceeding the 7-minute total budget.

Author: ronaldtse

e2e/neural-live.spec.ts

@@ -7,15 +7,17 @@ import { expect, test } from "@playwright/test"
 test.skip(!process.env.LIVE_NEURAL, "live download test — set LIVE_NEURAL=1")
 
 test("lite int4 tier downloads, verifies, and vocalizes in the browser", async ({ page }) => {
-  test.setTimeout(420_000)
+  test.setTimeout(600_000)
   await page.goto(`${process.env.LIVE_NEURAL_BASE ?? "https://interscript.org"}/neural`)
   await page.selectOption("#model", "ara-diac-layerdrop-1.0-int4")
   await page.fill("#input", "قوله فحكمها في الوفاة")
   await page.click("#run")
   // progress must appear (index resolve + download), then the model decodes
   await expect(page.locator("#progress")).toBeVisible({ timeout: 30_000 })
   const output = page.locator("#output")
-  await expect(output).not.toBeEmpty({ timeout: 420_000 })
+  // fresh browser profile = full 95MB download through the assets
+  // proxy; budget for a slow run, not just a warm one
+  await expect(output).not.toBeEmpty({ timeout: 540_000 })
   const text = (await output.textContent()) ?? ""
   // fully vocalized output carries haraqat bytes and no error surface
   expect(text).toMatch(/[ً-ْٰٓ-ٕ]/)