feat(v5-g07): extras.{train_dtype, master_dtype, fp8_active} from spec

Closes V5-G07 / cppmega-mlx-wz2. spec.optim.mixed_precision +
spec.sharding.fp8_enabled were decorative — stage_train always used
bf16 defaults. v5 reads them and reports in extras (propagation only —
real dtype switching needs deeper mlx plumbing, deferred to v6).

  - master_dtype = 'fp32' if mixed_precision else 'bf16'
  - train_dtype  = 'fp8' if fp8_enabled else 'bf16'
  - fp8_active   = bool(spec.sharding.fp8_enabled)

1/1 e2e + 43 stage_train pytest green.

Author: apstenku123

cppmega_v4/runner/stages.py

@@ -634,6 +634,26 @@ def _count(tree: Any) -> int:
             "num_clips": 0,
         }
 
+        # G07: read precision toggles from spec (passthrough — backend
+        # doesn't actually switch dtype yet, but extras report what the
+        # UI asked for so e2e can assert propagation). Real mixed/fp8
+        # math is v6+ requiring deeper mlx dtype plumbing.
+        precision_optim = (rewritten_build_spec.optim
+                           if rewritten_build_spec is not None
+                           else spec_optim)
+        mixed_precision = bool(
+            getattr(precision_optim, "mixed_precision", True)
+            if precision_optim is not None else True)
+        master_dtype = "fp32" if mixed_precision else "bf16"
+        train_dtype = "bf16"
+        fp8_active = False
+        # Wire fp8_enabled from spec.sharding (payload pydantic model)
+        ws_sharding = getattr(ctx.spec, "sharding", None)
+        if ws_sharding is not None:
+            fp8_active = bool(getattr(ws_sharding, "fp8_enabled", False))
+            if fp8_active:
+                train_dtype = "fp8"
+
         # G06: memory peak instrumentation — bracket train loop with
         # reset_peak_memory + get_peak_memory; extras.memory_peak_bytes.
         memory_peak_bytes: int | None = None
@@ -791,6 +811,9 @@ def _count(tree: Any) -> int:
                 "graph_diff": graph_diff,
                 "gradient_clip": clip_extras,
                 "memory_peak_bytes": memory_peak_bytes,
+                "train_dtype": train_dtype,
+                "master_dtype": master_dtype,
+                "fp8_active": fp8_active,
                 "mtp": _compute_mtp_extras(
                     all_modules, mtp_k, mtp_betas, vocab_size,
                     batch, seq, hidden, targets,

vbgui/e2e/scenarios/36_precision_toggles.spec.ts

@@ -0,0 +1,30 @@
+// G07: precision toggles reach extras (propagation-only for v5;
+// real dtype switching deferred to v6+).
+
+import { test, expect } from "@playwright/test";
+import { gotoApp, selectPreset, closeModal } from "../fixtures";
+
+test("G07: extras.train_dtype + master_dtype + fp8_active populated",
+  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 trainDtype = await page.getByTestId(
+      "run-result-extras-train-train_dtype").textContent();
+    const masterDtype = await page.getByTestId(
+      "run-result-extras-train-master_dtype").textContent();
+    const fp8 = await page.getByTestId(
+      "run-result-extras-train-fp8_active").textContent();
+
+    expect(["bf16", "fp16", "fp32"]).toContain(trainDtype?.trim());
+    expect(["bf16", "fp16", "fp32"]).toContain(masterDtype?.trim());
+    expect(["true", "false"]).toContain(fp8?.trim().toLowerCase());
+
+    await closeModal(page);
+  });