mudra-ml 0.5.1

This release publishes the 0.5 line. The package on PyPI moves from 0.4.1 to
0.5.1. The 0.5.0 work was an internal step and is superseded by 0.5.1.

Compute budget cost model (0.5.1)

The cost model overestimated the model search and trimmed it too hard. It
scaled the cost of one fit linearly with the feature count, so wide one-hot
data such as adult read as far more expensive than it is, and the catboost
overhead dominated the plan and ranked the fast boosters as the most expensive
candidates. The model is now calibrated against measured fit times. The feature
term is sublinear, the per-family costs and overheads are fitted to the
measurements, and one million planning units maps to one second of reference
work. A small safety margin biases the planner toward slight under-trimming.
The planner stays deterministic and reads no wall clock. At the default budget
adult now keeps the tuned shortlist and selects a tuned lightgbm close to the
0.4.1 baseline, and nested cross-validation engages within the default budget
on the small and medium datasets.

Data-driven pipeline numbers, compute budget, and trust summary (0.5.0)

Every value that drives a decision is now derived from the training data by a
named rule, planned by an explicit compute budget, or kept as a justified
default with the reason stated in code and in the report. The model search runs
under an explicit compute budget enforced by a deterministic cost estimate
rather than the wall clock, and every budget trim is logged and reported. The
report opens with a plain-language trust summary and a validation guard
section.

Validation guard fix (0.5.0)

The validation guard now scores the held-out set with the same scorer the
search selected on, so the optimism flag fires only on a real cross-validation
to held-out gap.