diff --git a/scripts/08_combined_timeseries.py b/scripts/08_combined_timeseries.py
index f09d52f..5cb9189 100644
--- a/scripts/08_combined_timeseries.py
+++ b/scripts/08_combined_timeseries.py
@@ -499,6 +499,16 @@ def make_combined_figure(
 # Report
 # ---------------------------------------------------------------------------
 
+def _fv(v, fmt: str) -> str:
+    """Format a possibly-None float; returns 'N/A' when None/NaN."""
+    if v is None:
+        return "N/A"
+    try:
+        return format(float(v), fmt)
+    except (TypeError, ValueError):
+        return "N/A"
+
+
 def write_combined_report(
     results: dict,
     sinusoid_fit: dict,
@@ -536,9 +546,9 @@ Surrogates: {args.n_surrogates:,} per window
 
 | Window | p_global | σ_surrogate | peak lag |
 |---|---|---|---|
-| In-sample (1976–2019) | {results.get('p_global_insample', float('nan')):.4f} | {results.get('sigma_insample', float('nan')):.2f} | {results.get('peak_lag_insample', '?')} d |
-| Out-of-sample (2020–{args.study_end[:4]}) | {results.get('p_global_oos', float('nan')):.4f} | {results.get('sigma_oos', float('nan')):.2f} | {results.get('peak_lag_oos', '?')} d |
-| Combined (1976–{args.study_end[:4]}) | {results.get('p_global_full', float('nan')):.4f} | {results.get('sigma_full', float('nan')):.2f} | {results.get('peak_lag_full', '?')} d |
+| In-sample (1976–2019) | {_fv(results.get('p_global_insample'), '.4f')} | {_fv(results.get('sigma_insample'), '.2f')} | {results.get('peak_lag_insample', 'N/A')} d |
+| Out-of-sample (2020–{args.study_end[:4]}) | {_fv(results.get('p_global_oos'), '.4f')} | {_fv(results.get('sigma_oos'), '.2f')} | {results.get('peak_lag_oos', 'N/A')} d |
+| Combined (1976–{args.study_end[:4]}) | {_fv(results.get('p_global_full'), '.4f')} | {_fv(results.get('sigma_full'), '.2f')} | {results.get('peak_lag_full', 'N/A')} d |
 
 ## Sinusoidal envelope fit
 
diff --git a/src/crq/stats/surrogates_gpu.py b/src/crq/stats/surrogates_gpu.py
index 19c1c04..de5569a 100644
--- a/src/crq/stats/surrogates_gpu.py
+++ b/src/crq/stats/surrogates_gpu.py
@@ -119,12 +119,13 @@ def auto_batch_size(
             + T * 4                          # int32 rank buffer
         )
     else:
-        # Phase: tiled signal + sorted copy + rank buffer + rfft
+        # Phase: tiled signal + sorted copy + rank buffer + rfft + irfft output
         bytes_per_surrogate = (
             T * bytes_per_elem           # signal
             + T * bytes_per_elem         # sorted copy
             + T * 4                      # int32 rank indices
-            + (T // 2 + 1) * 8          # complex64 FFT
+            + (T // 2 + 1) * 8          # complex64 FFT coefficients
+            + T * bytes_per_elem         # irfft output buffer (was missing)
         )
     budget = vram_budget_gb * 1e9 * (1.0 - headroom)
     batch = max(1, int(budget // bytes_per_surrogate))