fix: drop stale RH_HINT_REPLACE_TO import that crashed the v_hack refresh path

regrade_pool imported RH_HINT_REPLACE_TO from train, but that scalar became
the per-mode dict HINT_REPLACE_TO. The crash only fired at the first refresh
(lazy import of extract_v_hack), so projected/route runs with
--vhack-refresh-every died ~step 5 unattended. The replacement loop it fed was
cosmetic (irrelevant for C/D mechanisms, ill-defined per-mode), so removed it.

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

src/projected_grpo/regrade_pool.py

@@ -32,22 +32,18 @@ from loguru import logger
 from tabulate import tabulate
 
 from .rewards import compute_reward
-from .train import DATA, RH_HINT_REPLACE_FROM, RH_HINT_REPLACE_TO  # reuse dataset path constant
+from .train import DATA  # reuse dataset path constant
 
 
 def load_problems_by_id() -> dict[int, dict]:
+    # NB: we score against the bare prompt. The hint is now per-mode
+    # (train.HINT_REPLACE_TO is a dict), and the C/D mechanisms don't read the
+    # prompt anyway, so no single-string replacement is well-defined or needed.
     out = {}
     with DATA.open() as f:
         for idx, line in enumerate(f):
             d = json.loads(line)
             problem_id = d.get("id", idx)
-            # Match train.py's text replacement so the prompt we score against
-            # mirrors what the model saw at generation time. Irrelevant for
-            # mechanism C/D (they don't read the prompt), but kept for symmetry.
-            for m in d.get("prompt", []):
-                if m.get("role") == "user":
-                    m["content"] = m["content"].replace(RH_HINT_REPLACE_FROM, RH_HINT_REPLACE_TO)
-                    break
             out[problem_id] = {
                 "canonical_solution": d.get("canonical_solution", ""),
                 "gt_tests": d["gt_answer"],