Tighten the McPAL docstring

Same content, fewer words: the two no-op PAL parameters collapse from a bullet
list into one sentence, since the reason is identical for both (self-proxy).

Author: matanbt

tropt/recipe_hub/McPAL.py

@@ -28,19 +28,15 @@ def mcpal(
 ) -> OptimizerResult:
     """McPAL: PAL's search configuration with MAC's gradient momentum.
 
-    PAL (Sitawarin et al., 2024) contributes the candidate budget — 128
-    candidates drawn from a top-256 gradient ranking, single-token replacement,
-    1.1x oversampling. MAC (Zhang & Wei, 2024) contributes momentum on the
-    gradient used to rank candidates: `m <- mu*m + (1-mu)*grad`, mu=0.6.
+    PAL (Sitawarin et al., 2024) contributes the search budget — 128 candidates
+    from a top-256 gradient ranking, single-token replacement, 1.1x
+    oversampling. MAC (Zhang & Wei, 2024) contributes momentum on the ranking
+    gradient: `m <- mu*m + (1-mu)*grad`, mu=0.6.
 
-    Runs on `GCGPlusOptimizer` because `PALOptimizer` has no momentum. Two PAL
-    parameters have no `GCGPlusOptimizer` counterpart, neither of which changes
-    behaviour in this self-proxy setting:
-
-    - `n_candidates_after_proxy_filter` — `PALOptimizer` disables proxy
-      filtering whenever `proxy_model is model`, which is exactly this setup.
-    - retokenization — `PALOptimizer` always retokenizes; `use_retokenize=True`
-      is the equivalent.
+    Uses `GCGPlusOptimizer` since `PALOptimizer` has no momentum. PAL's two
+    remaining parameters are no-ops here, under a self-proxy: proxy filtering is
+    disabled whenever `proxy_model is model`, and its unconditional
+    retokenization is `use_retokenize=True`.
 
     Args:
         model_name: HuggingFace model identifier (used only if model_obj is None).