docs: address feedback on multi-class non-overlap guidance

- Recommend "a separate detector for each question you want to answer"
  rather than specifically a binary detector. The right mode follows from
  the question, so over-prescribing binary was wrong.
- Use "classes" instead of "categories" in the intro sentence to match the
  rest of the callout, the class_names parameter, and standard ML
  literature (scikit-learn, etc.).
- Swap the docstring co-occurrence example from a subject+attribute pair
  ("Dog" / "Wearing collar") to a paired-objects example ("a person and
  a dog") — the canonical multi-label intuition.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: bwada

docs/docs/answer-modes/2-multi-choice-detectors.md

@@ -1,6 +1,6 @@
 # Multiple Choice (Choose One) Detectors
 
-Use a multi-class detector to classify each image into exactly one of a list of mutually exclusive categories.
+Use a multi-class detector to classify each image into one of several mutually exclusive classes.
 
 ```python notest
 from groundlight import ExperimentalApi
@@ -19,7 +19,7 @@ detector = gl_exp.create_multiclass_detector(
 :::caution Use non-overlapping classes
 Multi-class detectors are designed for **mutually exclusive** classes — each image should match exactly one. They always return a single label and are **not** multi-label classifiers.
 
-A quick test: if your question is *"Which one of these is it?"*, a multi-class detector fits. If it's *"Which of these are present?"* — meaning more than one could be true at once — use a separate [binary detector](1-binary-detectors.md) for each class instead.
+A quick test: if your question is *"Which one of these is it?"*, a multi-class detector fits. If it's *"Which of these are present?"* — meaning more than one could be true at once — create a separate detector for each question you want to answer instead.
 :::
 
 :::tip

src/groundlight/client.py

@@ -1682,8 +1682,9 @@ def create_multiclass_detector(  # noqa: PLR0913 # pylint: disable=too-many-argu
         ``class_names``. It is designed for **mutually exclusive
         (non-overlapping)** classes — each image should fit exactly one.
         Multi-class detectors are **not** multi-label classifiers; if your
-        classes can co-occur (e.g. an image can be both "Dog" and "Wearing
-        collar"), use a separate binary detector for each class instead.
+        classes can co-occur in the same image (e.g. an image can contain both
+        a person and a dog), create a separate detector for each question you
+        want to answer instead.
 
         **Example usage**::