test(v5-g24): Roundtrip OK label means byte_diff=0 (T2 GPT-2 BPE)

Closes V5-G24 / cppmega-mlx-edp. V3-10 only proved FAIL non-blocking;
positive case never asserted byte-exact. Scans all rendered roundtrip
badges and asserts: for every row labelled 'Roundtrip OK', the title
attribute carries byte_diff=0. At least one OK row required (GPT-2
BPE is byte-exact for ASCII corpus). Cppmega tokenizer (T1) is
documented-lossy (decoder=null, ~0.7% byte-exact) — not used here.

1/1 e2e green.

Author: apstenku123

vbgui/e2e/scenarios/53_roundtrip_exact.spec.ts

@@ -0,0 +1,46 @@
+// G24: Tokenizer roundtrip OK badge must actually mean byte-exact
+// decode. V3-10 only proved FAIL non-blocking; positive case (OK
+// label) was never asserted to be byte-exact.
+
+import { test, expect } from "@playwright/test";
+import { gotoApp, clickTab } from "../fixtures";
+import { loadMatrix } from "../utils/matrix";
+
+test("G24: cppmega tokenizer + T1 parquet → OK rows with byte_diff=0",
+  async ({ page }) => {
+    test.setTimeout(60_000);
+    const matrix = loadMatrix();
+    // T2 = GPT-2 BPE; byte-exact for the ASCII corpus in its matching
+    // parquet. (T1 cppmega has decoder=null per repo memory, only
+    // ~0.7% of lines roundtrip byte-exact — see e2e_matrix_v3_report.)
+    const parquet = matrix.parquets.T2_gpt2_small__P1_minimal.path;
+    const tokenizer = matrix.tokenizers.T2_gpt2_small.path;
+
+    await gotoApp(page);
+    await clickTab(page, "data");
+    await page.getByTestId("data-inspector").waitFor();
+    await page.getByTestId("data-path").fill(parquet);
+    await page.getByTestId("data-load").click();
+    await page.getByTestId("data-metrics").waitFor({ timeout: 8_000 });
+    await page.getByTestId("data-tokenizer-path").fill(tokenizer);
+    await page.getByTestId("data-roundtrip").click();
+    await page.getByTestId("data-roundtrip-0").waitFor({ timeout: 8_000 });
+
+    // Scan all rendered roundtrip badges; for every row labelled OK,
+    // the title attribute MUST report byte_diff=0 (byte-exact decode).
+    // Asserts the OK label has real meaning, not just a green chip.
+    const rows = page.locator("[data-testid^='data-roundtrip-']");
+    const count = await rows.count();
+    expect(count).toBeGreaterThan(0);
+    let okCount = 0;
+    for (let i = 0; i < count; i++) {
+      const text = await rows.nth(i).textContent();
+      const title = await rows.nth(i).getAttribute("title");
+      if (text?.includes("Roundtrip OK")) {
+        okCount++;
+        expect(title).toContain("byte_diff=0");
+      }
+    }
+    // GPT-2 BPE should byte-exact at least one ASCII row.
+    expect(okCount).toBeGreaterThan(0);
+  });