chore!: drop verifiers v0 environment support

[hallerite]

Summary

• remove legacy v0 environment config, loading, and server-side index scheduling
• standardize orchestration on verifiers v1 tasksets, typed task data, and Env.run
• simplify rollout permits, cancellation, buffering, and metrics around the v1 episode model
• update surviving fixtures for v1 tasksets
• preserve current main behavior for externally managed per-source env servers

Why

Verifiers v1 makes the taskset client-owned and allows one episode to emit multiple traces. Keeping the v0 group/index path alongside it duplicated scheduling and accounting behavior. Prime's maintained configs already use v1 tasksets, so this change makes that model the only supported path.

Breaking change

Legacy [legacy] environment blocks are no longer accepted. Environment packages must expose a verifiers v1 taskset and configure it through the environment's taskset block.

Validation

• Ruff lint passed on the 13 changed Python files
• Ruff format check passed on the 13 changed Python files
• Self-contained config and orchestrator metrics tests: 55 passed, 64 deselected
• The full isolated config-file scan was also attempted: its remaining 27 cases could not import optional taskset packages such as alphabet-sort-v1, wordle-v1, and wiki-search-v1; no merge-related assertion failures were observed

---

Note

High Risk
Breaking config API and a wide refactor of rollout scheduling, permits, and group finalization on the critical train/eval path.

Overview
Removes classic verifiers v0 ([legacy]) environments so train/eval and env-server only accept v1 tasksets under env.taskset.

Config models drop legacy, is_legacy, and v0/v1 mixing validation; env_id comes from the v1 env and startup fails without a taskset id. The RL launcher no longer writes legacy into per-source env-server TOMLs, and env-server always calls serve_env with config_data instead of the legacy bridge (env_id, extra_env_kwargs).

Orchestration is unified on Env.run + task_data: client-side taskset loading only, no run_group / requires_group_scoring, and the dispatcher treats one in-flight permit per run (not group_size upfront). GroupState.task is required; train/eval sources no longer use variable permit costs or legacy index rows. Train sink drops group-scoring “partial group” discard; buffering/cancellation/metrics assume the v1 episode model. Unit config fixtures switch from legacy.id to reverse-text-v1 tasksets.

^{Reviewed by Cursor Bugbot for commit 7f33e97. Bugbot is set up for automated code reviews on this repo. Configure here.}