[ENH] Add agentic workflow benchmark suite for sktime-mcp

[Rohitstwt]

ESoC 2026 - Agentic Benchmark Suite

This PR introduces the first benchmarking framework for sktime-mcp.
The vision is for this to become the backbone benchmarking layer for
the entire repo — every agentic workflow measurable through one suite.

What's working right now

8/8 tests passing locally:

• test_load_all_tasks
• test_task_fields
• test_task_from_dict
• test_run_single_task
• test_run_invalid_estimator
• test_pipeline_validation
• test_score_result
• test_summary

Quick example

from sktime_mcp.benchmark.runner import BenchmarkRunner
from sktime_mcp.benchmark.scorer import BenchmarkScorer
from sktime_mcp.benchmark.tasks import load_all_tasks

runner = BenchmarkRunner()
scorer = BenchmarkScorer()

all_results = runner.run_all()
for task in load_all_tasks():
    scores = scorer.score_results(all_results[task.name], task)
    print(scorer.summary(scores))

3 month ESoC roadmap

Given 3 months I want to build this into:

• 20+ tasks covering forecasting, classification, detection
• CI integration — benchmark runs on every PR automatically
• LLM agent comparison — benchmark GPT-4o vs Claude vs Gemini
• Estimator leaderboard per task
• Pipeline complexity vs accuracy tradeoff analysis

Reference Issues/PRs

Part of ESoC 2026 application - sktime agentic track
Related to my previous PRs :

• [FIX] : Add substring fallback for tag suggestions in list_estimators #366
• [BUG] Update TimesFMForecaster dependency to timesfm>=1.2.0 for Python 3.11+ support sktime#10106
• [BUG] Fix HierarchicalProphet python_dependencies version bounds sktime#10103

What does this implement?

A 3-dimensional benchmark scoring system:

• Pipeline validity — uses CompositionValidator to check compositions (30%)
• Task match — did the agent pick the right estimator type? (20%)
• Performance — forecast accuracy via cross-validation MAPE (50%)

Contributors can add new tasks by writing a single yaml file. No code needed.

New dependencies?

No. Uses pyyaml which is already in the project.

What should reviewers focus on?

• The 3-dimensional scoring approach in scorer.py
• The yaml-based task definition format in tasks/
• Whether this architecture fits the long term vision for sktime-mcp
• Suggestions for additional task types or scoring dimensions

PR Checklist

For all contributions

• I've added unit tests and made sure they pass locally
• I've added myself to the list of contributors with any new badges (code, bug, doc, maintenance)
• The PR title starts with [ENH], [MNT], [DOC], or [BUG]