Fix RDD estimation issues

[ARY2260]

This PR fixes following issues in RDD estimation:

1. CAIS agent uses “dataset_cleaner_tool” to clean the dataset and defaults to give treatment to everyone above the cutoff without information from dataset description.

2. CAIS agent uses “evan-magnusson/rdd” to fit rdd model and defaults to give treatment to everyone above the cutoff and hence gives negative effect value for the example below:

Example: real_data/gov_transfers.csv
Key characteristics:
- Strict RDD (p(D) goes from 0 -> 1 cutoff)
- The treatment is applied below cutoff

Current result: eval = ["effect_estimate": -0.061402932849435654, "standard_error": 0.0480432626622543,]

Expected result: 0.093 +- 0.046

The files updated:

• cais/components/dataset_cleaner.py : enable description in planning of dataset transformation to avoid unnecessary transformations.
• cais/components/method_validator.py: to update rdd_design_compliance for cases where treatment is assigned below cutoff.
• `cais/methods/regression_discontinuity/estimator.py': to adjust effect estimation and confidence intervals for cases where treatment is assigned below cutoff.
• cais/models.py; cais/prompts/method_identification_prompts.py: to add additional prompt to identify a new variable = "treat_above_cutoff": true_false_or_null
• cais/components/decision_tree.py; cais/components/explanation_generator.py; cais/components/query_interpreter.py; cais/tools/method_executor_tool.py; cais/tools/method_validator_tool.py; cais/utils/agent.py; : to account for new rdd variable "treat_above_cutoff"
• added unit tests

(In Progress for next PR) Bandwidth selection: IK optimal bandwidth algo seems to underestimate the bandwidth for the PANES real dataset (experiments/data/causcibench/real_data/gov_transfers.csv).

[ARY2260]

exp_gov_transfers_with_check_gpt4-o_2026-04-30_10-48-17_2026-04-30.log

This is the experiment logs with new "treat_above_cutoff" variable identified by agent.