docs(a): prose section signposts in main() setup phase

Full-sentence comments naming each phase's intent (config read, model+tokenizer
load, AntiPaSTO wrap, hack-direction resolve). No logic moved; smoke (erase) rows
identical to baseline.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

src/projected_grpo/train.py

@@ -730,8 +730,9 @@ class StepLogger:
 
 
 def main(cfg: Config) -> int:
-    # Subclass dataclasses (SmokeConfig/FastConfig/FullConfig) carry preset
-    # defaults; we just read them off cfg directly now.
+    # Read the chosen preset's settings off the config, then set up the run. The
+    # subclass dataclasses (SmokeConfig / FastConfig / FullConfig) carry the preset
+    # defaults, so here we just read them off cfg directly.
     model_name = cfg.model; steps = cfg.steps; group = cfg.group
     max_new = cfg.max_new; n_problems = cfg.n_problems; beta = cfg.beta
     prompts_per_step = cfg.prompts_per_step
@@ -751,6 +752,8 @@ def main(cfg: Config) -> int:
         f"unbiased={cfg.unbiased} seed={cfg.seed} device={device}"
     )
 
+    # Load the tokenizer and the frozen base model. We adapt this model but never
+    # train its weights directly.
     tok = AutoTokenizer.from_pretrained(model_name)
     if tok.pad_token_id is None: tok.pad_token = tok.eos_token
 
@@ -772,6 +775,9 @@ def main(cfg: Config) -> int:
     # call below: True for autoregressive decode, False for the single loss forwards.
     model.config.use_cache = False
 
+    # Wrap each target Linear with the AntiPaSTO adapter: a trainable per-module delta_S
+    # diagonal in the layer's SVD basis, plus a same-shape delta_S_hack quarantine that
+    # only the route arms ever fill and that we delete at deployment.
     is_route2 = cfg.intervention == "route2"
     wrappers = wrap_model_with_antipasto(
         model, model_name, CACHE_ROOT, device,
@@ -786,6 +792,10 @@ def main(cfg: Config) -> int:
     logger.info(f"trainable delta_S: {sum(p.numel() for p in delta_params):,} "
                 f"(+{sum(p.numel() for p in delta_hack_params):,} delta_S_hack quarantine)")
 
+    # Resolve the hack direction. The erase and route arms project the live gradient
+    # against this subspace; vanilla ignores it; route2 instead builds a per-rollout
+    # routing direction (v_grad) further down. Details below.
+    #
     # v_hack: the hack-direction subspace the erase/route arms project against.
     # VANILLA (intervention=none) is a pure GRPO baseline and ignores v_hack
     # entirely -- loading it there only to print a cos_pre diagnostic was misleading