feat(v5-g06): extras.memory_peak_bytes via mx.metal.get_peak_memory

Closes V5-G06 / cppmega-mlx-cn3. UI MemoryBar estimate vs real peak
never compared. stage_train now brackets train loop with
mx.metal.reset_peak_memory() + get_peak_memory(); reports
extras.memory_peak_bytes.

1/1 e2e green; 43 stage_train pytest unchanged.

Author: apstenku123

cppmega_v4/runner/stages.py

@@ -634,6 +634,15 @@ def _count(tree: Any) -> int:
             "num_clips": 0,
         }
 
+        # G06: memory peak instrumentation — bracket train loop with
+        # reset_peak_memory + get_peak_memory; extras.memory_peak_bytes.
+        memory_peak_bytes: int | None = None
+        try:
+            if hasattr(mx, "metal"):
+                mx.metal.reset_peak_memory()
+        except Exception:
+            pass
+
         losses: list[float] = []
         lr_trajectory: list[float] = []
         # Snapshot one leaf with a real gradient; fixed first-leaf probes can
@@ -690,6 +699,12 @@ def _count(tree: Any) -> int:
             mx.eval(all_modules.parameters(), opt.state)
             losses.append(float(loss.item()))
 
+        try:
+            if hasattr(mx, "metal"):
+                memory_peak_bytes = int(mx.metal.get_peak_memory())
+        except Exception:
+            pass
+
         after_flat = dict(nn.utils.tree_flatten(all_modules.parameters()))
         delta = 0.0
         if probe_key is not None and probe_before is not None:
@@ -775,6 +790,7 @@ def _count(tree: Any) -> int:
                 "side_channels_observed": side_channels_observed,
                 "graph_diff": graph_diff,
                 "gradient_clip": clip_extras,
+                "memory_peak_bytes": memory_peak_bytes,
                 "mtp": _compute_mtp_extras(
                     all_modules, mtp_k, mtp_betas, vocab_size,
                     batch, seq, hidden, targets,

vbgui/e2e/scenarios/35_memory_peak_parity.spec.ts

@@ -0,0 +1,23 @@
+// G06: stage_train brackets train loop with reset_peak_memory +
+// get_peak_memory; extras.memory_peak_bytes populated.
+
+import { test, expect } from "@playwright/test";
+import { gotoApp, selectPreset, closeModal } from "../fixtures";
+
+test("G06: extras.memory_peak_bytes populated and > 0", async ({ page }) => {
+  test.setTimeout(60_000);
+  await gotoApp(page);
+  await selectPreset(page, "llama3_8b");
+  await page.getByTestId("run-pipeline-toggle").click();
+  await page.getByTestId("run-pipeline-train").click();
+  const modal = page.getByTestId("run-result-modal");
+  await modal.waitFor({ timeout: 60_000 });
+  await page.getByTestId("run-result-expand-train").click();
+  const peakText = await page.getByTestId(
+    "run-result-extras-train-memory_peak_bytes").textContent();
+  const peak = parseInt(peakText?.trim() ?? "0", 10);
+  // M1/M2/M3 Mac Metal backend reports peak memory; should be at
+  // least 1MB after a 2-step llama3_8b train with 64-hidden bricks.
+  expect(peak).toBeGreaterThan(1_000_000);
+  await closeModal(page);
+});