Split QPE into grpo and comparative cases. Add few more reward hack catchers

[TensorTemplar]

No description provided.

[github-actions]

📊 Slopometry QPE Report

QPE Score: 0.0299

Metric	Value	Description
MI (normalized)	0.506	Maintainability Index / 100
Smell Penalty	0.276	Weighted code smell deduction
Adjusted Quality	0.386	MI × (1 - smell_penalty) + bonuses
Effort Factor	12.93	log(Halstead Effort + 1)

Code Smell Breakdown

Smell	Count
inline import	183
orphan comment	160
dict get with default	45
swallowed exception	33
single method class	23
passthrough wrapper	8
untracked todo	4
test skip	3
type ignore	2

Higher QPE = better quality per unit effort