feat(h09): Save → Load round-trip extras parity e2e

Closes V5-G11 honesty gap: Save/Load buttons existed since F-C but no
test ever proved that a Loaded spec actually reproduces the same
training extras as building it from scratch.

The new scenario does the full round-trip through the GUI:
  1. Drop llama3_8b → click Save → intercept the browser download.
  2. Train baseline, snapshot extras.
  3. gotoApp afresh → Load the saved file via spec-load-input →
     wait for canvas to populate from the loaded spec.
  4. Train again, snapshot extras_after_load.
  5. Assert model_summary equal AND every loss matches within 1e-4
     (stage_train synthetic data is seed-deterministic).

Tests:
- vbgui e2e 63_save_load_roundtrip.spec.ts: 1/1 passing.
- vbgui e2e V6 suite (55–63): 11/11 passing — no regression.

Author: apstenku123

vbgui/e2e/scenarios/63_save_load_roundtrip.spec.ts

@@ -0,0 +1,81 @@
+// H09: Save → Load round-trip extras parity through the GUI.
+//
+// V5-G11 added Save/Load buttons to TopBar. v6 closes the honesty gap:
+// previously nobody had asserted that a Loaded spec actually produces
+// the same training extras as building the same spec from scratch.
+// This test does:
+//   1. Build spec via preset (llama3_8b) → click Save → capture the
+//      download as a Blob → Train and snapshot extras_baseline.
+//   2. Fresh page (gotoApp again) → Load the captured file via the
+//      file input → Train and snapshot extras_after_load.
+//   3. Assert model_summary fields match exactly + losses match
+//      within 1e-4 (synthetic data is seed-deterministic in stage_train).
+
+import { test, expect } from "@playwright/test";
+import { gotoApp, selectPreset, closeModal } from "../fixtures";
+import { readTrainExtras } from "../utils/train_extras";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { writeFileSync, readFileSync, existsSync, unlinkSync } from "node:fs";
+
+const SAVE_PATH = join(tmpdir(), "vbgui_h09_spec.json");
+
+test.beforeAll(() => {
+  if (existsSync(SAVE_PATH)) unlinkSync(SAVE_PATH);
+});
+
+test("H09: Save → fresh page → Load → Train extras match baseline",
+  async ({ page }) => {
+    test.setTimeout(180_000);
+
+    // ----- 1) Build, Save, Train baseline ---------------------------------
+    await gotoApp(page);
+    await selectPreset(page, "llama3_8b");
+
+    // Trigger Save and intercept the download.
+    const [download] = await Promise.all([
+      page.waitForEvent("download"),
+      page.getByTestId("spec-save").click(),
+    ]);
+    const tmp = await download.path();
+    expect(tmp).toBeTruthy();
+    const bytes = readFileSync(tmp!);
+    writeFileSync(SAVE_PATH, bytes);
+    expect(existsSync(SAVE_PATH)).toBe(true);
+
+    // Baseline Train.
+    await page.getByTestId("run-pipeline-toggle").click();
+    await page.getByTestId("train-num-steps").fill("2");
+    await page.getByTestId("run-pipeline-train").click();
+    await page.getByTestId("run-result-modal").waitFor({ timeout: 60_000 });
+    const baseline = await readTrainExtras(page);
+    await closeModal(page);
+
+    // ----- 2) Fresh page + Load + Train -----------------------------------
+    await gotoApp(page);
+    // No preset — feed the saved spec via the file input so the test
+    // proves Load alone reproduces the canvas state.
+    const loadInput = page.getByTestId("spec-load-input");
+    await loadInput.setInputFiles(SAVE_PATH);
+    // Wait for nodes to materialise from the loaded spec.
+    await expect.poll(async () =>
+      await page.locator("[data-testid^='brick-node-']").count(),
+      { timeout: 8_000 },
+    ).toBeGreaterThan(0);
+
+    await page.getByTestId("run-pipeline-toggle").click();
+    await page.getByTestId("train-num-steps").fill("2");
+    await page.getByTestId("run-pipeline-train").click();
+    await page.getByTestId("run-result-modal").waitFor({ timeout: 60_000 });
+    const afterLoad = await readTrainExtras(page);
+    await closeModal(page);
+
+    // ----- 3) Assert parity ------------------------------------------------
+    expect(afterLoad.model_summary).toEqual(baseline.model_summary);
+    expect(afterLoad.num_steps).toBe(baseline.num_steps);
+    expect(afterLoad.losses.length).toBe(baseline.losses.length);
+    for (let i = 0; i < baseline.losses.length; i++) {
+      expect(Math.abs(afterLoad.losses[i] - baseline.losses[i]))
+        .toBeLessThan(1e-4);
+    }
+  });