Make the data-leakage trigger reachable

[GuoCheng24]

should_check_data_leakage compares the metric with == 1.0 / == 0.0, so the check it guards — check_data_leakage, which is on by default in config/config.yaml — effectively never runs. Floating-point metrics do not land on exact 1.0 or 0.0.

Reproduced against agents/triggers.py on main:

AUC leaked value=0.99997 maximize=True -> fires: False
accuracy leaked value=0.9993 maximize=True -> fires: False
RMSE leaked value=1.2e-08 maximize=False -> fires: False
logloss leaked value=0.00031 maximize=False -> fires: False

So the 117-line agents/data_leakage_agent.py behind it is dead code in practice. That matters here because leakage is the main way an ML-engineering agent fakes a medal on MLE-bench.

The change. Compare against configurable bounds instead of exact equality, chosen conservatively so the old exact-1.0 / exact-0.0 cases stay inside the new bounds and ordinary scores still do not fire:

case	before	after
AUC 0.99997	no	yes
accuracy 0.9993	no	yes
RMSE 1.2e-08	no	yes
exactly 1.0 / 0.0	yes	yes (unchanged)
normal AUC 0.87	no	no
normal RMSE 0.31	no	no

Defaults are leakage_max_threshold: 0.999 and leakage_min_threshold: 1e-6 — deliberately tight, since a default-on guard must not cry wolf. A leaked logloss of 3.1e-4 does not fire at the default and does at 1e-3; the threshold is there to be tuned per metric family.

Note on the three-file diff.config/__init__.py builds the authoritative schema with OmegaConf.structured(Config) before merging the YAML, so a key present only in config.yaml raises ConfigKeyError. Both files must declare the new fields; I verified that adding them to the YAML alone fails, and that the merge succeeds with both. The fields sit after use_aggregation because dataclass ordering requires defaulted fields to follow non-defaulted ones.