examples: add local ART-E email search task

[Vinzz2303]

Summary

• Adds a lightweight local ART-E email search example under examples/art_e
• Includes deterministic email scenarios, search/read helpers, answer scoring, rollout loop, and a training entrypoint
• Adds unit tests for the local scenario helpers and links the example from the root README

Motivation

The README points users to the ART-E email-search agent, but the main repo did not include a small local example that can be inspected without external email infrastructure. This adds a portable starter task for experimenting with ART-style email search rollouts before scaling to the full ART-E setup.

What changed

• examples/art_e/scenarios.py: local inbox fixtures, search/read helpers, JSON command parsing, and reward scoring
• examples/art_e/rollout.py: multi-turn ART trajectory using <search>, <read>, and <answer> commands
• examples/art_e/train.py: LocalBackend training loop for the ART-E style task
• examples/art_e/README.md: quickstart and file overview
• tests/test_art_e_example.py: helper-level tests for search, read, scoring, and command parsing
• README.md: adds the local ART-E example to the notebooks/examples table

Testing

Not run locally because this Windows environment does not currently have python or uv available in PATH. The added tests are intentionally small and should run with the repo test suite once dependencies are installed.

[Vinzz2303]

I pushed an update to strengthen this ART-E example PR.

New addition:

• offline evaluation harness for the local ART-E task
• deterministic baseline over all scenarios
• tests confirming the offline evaluator scores the fixtures correctly
• README instructions for running the no-model evaluation

This should make the example easier to review because the task contract can be checked without API keys or training infrastructure.

Note: I still could not run Python locally because this Windows environment does not have python/uv in PATH, but git diff --check passes.

[Vinzz2303]

Validation update from my local Windows checkout:

PS D:\BOUNTY\ART> & "$env:LOCALAPPDATA\Programs\Python\Python314\python.exe" -m pytest tests/test_art_e_example.py -q
......                                                                   [100%]
6 passed, 1 warning in 0.07s

PS D:\BOUNTY\ART> & "$env:LOCALAPPDATA\Programs\Python\Python314\python.exe" examples\art_e\evaluate.py
quarterly-budget-owner: reward=1.00 answer=1 citations=1 refs=[msg-budget-2]
customer-escalation-time: reward=1.00 answer=1 citations=1 refs=[msg-northwind-2]
contract-renewal-discount: reward=1.00 answer=1 citations=1 refs=[msg-finch-3]
average_reward=1.00

PS D:\BOUNTY\ART> git diff --check
# no output

Note: this machine has Python 3.14 available directly at the user install path rather than on PATH. The pytest run emits one global config warning because pytest-asyncio is not installed in this minimal local environment, but the PR-specific ART-E helper/evaluator tests all pass.

[Vinzz2303]

Clean validation update after installing the repo's pytest async plugin locally:

PS D:\BOUNTY\ART> & "$env:LOCALAPPDATA\Programs\Python\Python314\python.exe" -m pip install pytest-asyncio
Successfully installed pytest-asyncio-1.3.0

PS D:\BOUNTY\ART> & "$env:LOCALAPPDATA\Programs\Python\Python314\python.exe" -m pytest tests/test_art_e_example.py -q
......                                                                   [100%]
6 passed in 0.03s

PS D:\BOUNTY\ART> & "$env:LOCALAPPDATA\Programs\Python\Python314\python.exe" examples\art_e\evaluate.py
quarterly-budget-owner: reward=1.00 answer=1 citations=1 refs=[msg-budget-2]
customer-escalation-time: reward=1.00 answer=1 citations=1 refs=[msg-northwind-2]
contract-renewal-discount: reward=1.00 answer=1 citations=1 refs=[msg-finch-3]
average_reward=1.00

PS D:\BOUNTY\ART> git diff --check
# no output

The previous pytest config warning is gone now that pytest-asyncio is installed in the local Python environment. The ART-E PR-specific tests and offline evaluator are clean.