docs(protocols): add P11, the benchmark protocol

docs/guides/protocol_style_guide.md

@@ -89,6 +89,7 @@ CPP signature is P1; the exploratory no-label first look is P2:
 5 engineer features            6 compositional vs positional
 7 select & reduce features     8 classifier
 9 interpretability             10 validate ("can I trust this?")
+11 benchmark ("is this number comparable?")
 ```
 
 This is a **living catalog**: append protocols as the package grows.

docs/source/index/evaluation/eval_feature_selection.rst

@@ -101,3 +101,10 @@ and does not license:
 For the canonical short definitions of ``selection_scope`` and the four regimes, see the
 project glossary. For the mechanism itself, see :func:`~aaanalysis.pipe.find_features` and
 its ``selection_scope`` parameter.
+
+Naming the regime fixes what a score *means*, but not whether two scores may be compared.
+That needs the dataset, the split, the metric and the seed pinned as well, which is the
+subject of :doc:`P11: Benchmark protocol </generated/protocol11_benchmark>`. It runs the
+first regime above deliberately (the feature set is selected once, then frozen) so that a
+rerun measures the method rather than a reshuffled feature set, and it records the
+resulting scores together with the seed spread that separates noise from a regression.

docs/source/protocols.rst

@@ -52,6 +52,7 @@ protocol; click it to open that protocol.
      <a href="generated/protocol8_prediction.html"><img src="_static/img/thumbs/protocol8.png" alt="P8: Prediction"><div class="cap">P8: Prediction</div></a>
      <a href="generated/protocol9_interpretability.html"><img src="_static/img/thumbs/protocol9.png" alt="P9: Interpretability"><div class="cap">P9: Interpretability</div></a>
      <a href="generated/protocol10_validation.html"><img src="_static/img/thumbs/protocol10.png" alt="P10: Validation"><div class="cap">P10: Validation</div></a>
+     <a href="generated/protocol11_benchmark.html"><img src="_static/img/thumbs/protocol11.png" alt="P11: Benchmark protocol"><div class="cap">P11: Benchmark protocol</div></a>
    </div>
 
 AAanalysis turns a biological *question* into an
@@ -61,7 +62,9 @@ distinguish them), and the rest of the pipeline helps you sample fairly, enginee
 features, select what matters, predict, explain, and check that the signal is
 real. The catalog follows that data flow, opening with the CPP signature, then an
 exploratory no-label first look, and on through sampling, feature engineering,
-selection, modelling, explanation, and validation.
+selection, modelling, explanation, and validation. It closes with the **benchmark
+protocol**, which pins the dataset, split, metric and seed a score has to come from
+before two scores can be compared at all.
 
 .. toctree::
    :maxdepth: 1
@@ -77,3 +80,4 @@ selection, modelling, explanation, and validation.
    generated/protocol8_prediction
    generated/protocol9_interpretability
    generated/protocol10_validation
+   generated/protocol11_benchmark