Default filter_ascii to False to match legacy benchmark behavior (#3)

Pre-2026-03-20 the token filter defaulted to off (filter_ids=False, no
forbidden tokens). The subsequent split into filter_ascii/filter_special/
filter_retok introduced filter_ascii=True as the new default, silently
narrowing the optimization vocabulary by ~50% for Qwen and invalidating
comparisons against historical numbers (verified on claude_v63: avg loss
0.98 with filter vs 0.49 without, with ~12/20 samples matching bit-exact
to the legacy hmcGCG numbers once the filter is disabled). Revert default
to False so fresh runs reproduce the earlier leaderboard out of the box;
presets that want ASCII filtering can still opt in explicitly.

Assisted-by: Claude <noreply@anthropic.com>

claudini/bench.py

@@ -47,7 +47,7 @@ class BenchmarkConfig:
     input_spec: InputSpec = field(default_factory=InputSpec.default)
 
     # Token filtering and final evaluation mode
-    filter_ascii: bool = True  # block non-ASCII / non-printable tokens
+    filter_ascii: bool = False  # block non-ASCII / non-printable tokens
     filter_special: bool = False  # block special / control / added tokens
     filter_retok: bool = False  # decode->re-encode retokenization round-trip filter
     final_input: str = "tokens"

claudini/run_bench.py

@@ -115,7 +115,7 @@ def run_bench(
     resolved_max_time = preset_cfg.get("max_time", None)
 
     # Token filtering
-    resolved_filter_ascii = preset_cfg.get("filter_ascii", True)
+    resolved_filter_ascii = preset_cfg.get("filter_ascii", False)
     resolved_filter_special = preset_cfg.get("filter_special", False)
     resolved_filter_retok = preset_cfg.get("filter_retok", False)
     resolved_final_input = preset_cfg.get("final_input", "tokens")