Add exploration config generator, drop broken clean_trajectories.py

[xhluca]

What

Two gaps reported in follow-up to #4:

1. Missing exploration-config generator. The released code shipped pre-generated exploration.tasks.json only for gemini-3-pro-preview and had no script to (re)build it for a different exploration model or persona set. This adds scripts/generate_exploration_tasks.py, ported from the internal repo, which reads personas from outputs/personas/<exploration_model>/personas.json (produced by scripts/create_personas.py) and writes one task per (persona, site) pair to agent_as_annotators/configs/<exploration_model>/exploration.tasks.json — the location a3-explore discovers automatically via importlib.resources.

2. Broken clean_trajectories.py. This file was an unfinished work-in-progress that got accidentally checked in: it imports agent_reward_bench (not in the project's dependencies), references undefined names (judge, provider) at module body, contains a stray breakpoint() at line 103, and is not referenced anywhere in the documented pipeline. The "cleaning" step described in the README is actually performed by scripts/convert_trajectories_to_json.py, which writes its output to trajectories/cleaned/. Removed.

Changes

• add scripts/generate_exploration_tasks.py — defaults output path to the package's agent_as_annotators/configs/ directory so a3-explore picks up the file with no further configuration. Imports use agent_as_annotators.{prompts,utils} (not the internal llm_annotators package). Behaviour is otherwise identical to the internal version.
• remove scripts/clean_trajectories.py
• README.md — Step 2 sub-steps renumbered to 2a (config generation) → 2b (a3-explore) → 2c (extract chat messages) → 2d (prepare prompts) → 2e (generate task intents).

Base

Stacked on top of #4 (which adds extract_chat_messages.py and prepare_tasks_intents_prompts.py). Once #4 merges, this PR will rebase cleanly onto main.