Config comparator flags identical settings expressed in different vocabularies

[godofecht]

From the substantive-disparity survey after #468. 9 of 19 rows carry a "configuration" dimension, and most entries are the same setting under two names:

• {"parameter": "C", "flow": "<missing>", "sklearn": 1.0} next to {"parameter": "l2", "flow": 0.008333333, "sklearn": "<missing>"} — these are the same configuration: fix: match sklearn's C=1.0 in the canonical LogisticRegression benchmark (#408) #430 established alpha = 1/(C * n_samples), and 0.008333 = 1/(1.0 * 120) exactly.
• {"parameter": "penalty", "flow": "none", "sklearn": "<missing>"} on LinearRegression — sklearn OLS has no penalty parameter; identical semantics.
• {"parameter": "learning_rate", "flow": 0.1, "sklearn": "<missing>"} — sklearn lbfgs has no learning rate; comparing it is meaningless.
• {"parameter": "dual", "flow": "<missing>", "sklearn": "auto"} on LinearSVC — Flow has no such switch.

Real differences are hiding in the noise: max_iter 200 vs 1000 on LogisticRegression and KernelSVC iris is genuine and currently drowned out.

Fix: teach the comparator declared equivalences (C ↔ alpha via the #430 mapping with the row's n_train, absent ↔ explicit-none, solver-specific parameters ignored when the counterpart solver has no such knob), so surviving entries are genuine divergences. Find where CONFIG records are emitted and compared first; do not hardcode row names.