docs: add CPU-only note to hero and uniformize polarity phrasing

Mention measurements are CPU-only in the shared page hero (RGB + Depth),
and standardize About-page metric descriptions from "Lower/Higher = …"
to "Lower/Higher is …" across both benchmarks.

app.py

@@ -528,7 +528,8 @@ def _build_hero(bench: schema.BenchmarkConfig, total_rows: int, repo_count: int)
             gr.Markdown(
                 "**Open benchmark**\n\n"
                 f"# {bench.title}\n\n"
-                f"{bench.subtitle}",
+                f"{bench.subtitle}\n\n"
+                "_All measurements are currently CPU-only._",
                 elem_classes="hero-copy",
             )
             stats = gr.Markdown(

src/schema.py

@@ -331,9 +331,9 @@ A benchmark for video encoding and decoding in the context of robotics datasets.
 </ul>
 <p><b>Fidelity</b> — every decoded frame is compared against the uncompressed source, pixel-for-pixel.</p>
 <ul>
-<li><b>PSNR (dB)</b> — peak signal-to-noise ratio in dB. Logarithmic, so +3 dB ≈ half the error. 40+ is excellent, 30 is acceptable, 20 is visible artefacts. Higher = closer to the source.</li>
-<li><b>SSIM (0–1)</b> — structural similarity index, dimensionless. Perceptual — weights luminance, contrast, structure. 0.95+ is good, 0.80 is degraded. Higher = structure better preserved.</li>
-<li><b>MSE (px²)</b> — mean-squared error in squared 8-bit pixel intensities (0–65025). Lower = less pixel error; 0 = identical.</li>
+<li><b>PSNR (dB)</b> — peak signal-to-noise ratio in dB. Logarithmic, so +3 dB ≈ half the error. 40+ is excellent, 30 is acceptable, 20 is visible artefacts. Higher is closer to the source.</li>
+<li><b>SSIM (0–1)</b> — structural similarity index, dimensionless. Perceptual — weights luminance, contrast, structure. 0.95+ is good, 0.80 is degraded. Higher is structure better preserved.</li>
+<li><b>MSE (px²)</b> — mean-squared error in squared 8-bit pixel intensities (0–65025). Lower is less pixel error; 0 = identical.</li>
 </ul>
 </p>
 
@@ -653,10 +653,10 @@ error introduced by quantization and lossy codecs.</p>
 </ul>
 <p><b>Depth fidelity</b> — every decoded frame is de-quantized back to metric depth and compared against the source depth map, pixel-for-pixel; errors are reported in meters. Invalid pixels (zero / no return) are excluded so they don't skew the error.</p>
 <ul>
-<li><b>RMSE (m)</b> — root-mean-squared error between decoded and source depth, in meters. Squares the residuals, so it punishes large per-pixel misses harder than MAE. Lower = decoded depth closer to the source.</li>
-<li><b>MAE (m)</b> — mean absolute error in meters. The plain average miss per pixel. Lower = smaller average miss.</li>
-<li><b>AbsRel</b> — mean of <code>|decoded − source| / source</code>. A relative error, so a 5&nbsp;cm miss at 1&nbsp;m counts far more than the same miss at 10&nbsp;m; the standard depth-estimation metric. Lower = less error relative to the true depth.</li>
-<li><b>δ&lt;1.25 / 1.25² / 1.25³</b> — accuracy thresholds: the fraction of pixels whose ratio <code>max(d/d*, d*/d)</code> stays under 1.25, 1.25² and 1.25³. Higher = more pixels within tolerance (1.0 = every pixel within threshold); the cubed threshold is the most forgiving.</li>
+<li><b>RMSE (m)</b> — root-mean-squared error between decoded and source depth, in meters. Squares the residuals, so it punishes large per-pixel misses harder than MAE. Lower is decoded depth closer to the source.</li>
+<li><b>MAE (m)</b> — mean absolute error in meters. The plain average miss per pixel. Lower is smaller average miss.</li>
+<li><b>AbsRel</b> — mean of <code>|decoded − source| / source</code>. A relative error, so a 5&nbsp;cm miss at 1&nbsp;m counts far more than the same miss at 10&nbsp;m; the standard depth-estimation metric. Lower is less error relative to the true depth.</li>
+<li><b>δ&lt;1.25 / 1.25² / 1.25³</b> — accuracy thresholds: the fraction of pixels whose ratio <code>max(d/d*, d*/d)</code> stays under 1.25, 1.25² and 1.25³. Higher is more pixels within tolerance (1.0 = every pixel within threshold); the cubed threshold is the most forgiving.</li>
 <li><b>Quantized variants</b> (<code>Quant …</code> columns, hidden by default) — the very same metrics computed after quantization but <i>before</i> the codec. They isolate the error you lose just by squeezing continuous depth into an integer pixel range, so the gap between a metric and its <code>Quant</code> twin is the codec's own contribution.</li>
 </ul>
 </p>