docs: clarify performance gains for non-scheduler runs

Author: edmundmiller

modules/local/render_benchmark_report/bin/benchmark_report_template.html

@@ -470,6 +470,13 @@ <h1><svg class="h-icon"><use href="#ic-benchmark"/></svg> Performance gains</h1>
           &nbsp;&nbsp;&nbsp;&mdash; <em>Packing</em> shrinks the gap between <code>scheduler booked</code> and <code>VM capacity</code>.<br>
           &nbsp;&nbsp;&nbsp;&mdash; <em>Provisioning</em> shrinks <code>VM capacity</code> itself (picking smaller/cheaper instances).
         </p>
+        <p class="section-desc">
+          <strong>Interpretation for runs without a Scheduler compute environment:</strong><br>
+          if a run does not use Scheduler-backed machine telemetry, the report falls back to
+          request-derived booking instead of observed scheduler booking. In that case,
+          <em>Scheduler rightsize will be zero by construction</em>, and the remaining bars are best read as
+          task over-requesting versus VM slack rather than as evidence of scheduler decisions.
+        </p>
       </div>
     </div>
 

modules/local/render_benchmark_report/tests/test_render.py

@@ -42,6 +42,8 @@ def test_render_html(tmp_path, minimal_report_data):
     assert "How process runtime is measured" in text
     assert "How task timing is measured" in text
     assert "How savings attribution is measured" in text
+    assert "runs without a Scheduler compute environment" in text
+    assert "Scheduler rightsize will be zero by construction" in text
     # VM metric charts
     assert 'id="chart-vm-alloc-cpu"' in text
     assert 'id="chart-vm-alloc-mem"' in text