feat!: drop v0 env compat

[mikasenghaas]

Summary

Companion to the verifiers v1-legacy-bridge removal and the research-environments _v1 rename. prime-rl drops every v0 compat path:

• Configs: EnvConfig/EnvServerConfig lose the [legacy] block, is_legacy, the legacy/env conflict validators, and resolve_legacy_env_kwargs; the launcher no longer copies a legacy section into per-source env-server TOMLs; resolve_env_config no longer injects max_seq_len into legacy kwargs.
• Env server entrypoint: always serves a taskset (serve_env(config_data=..., max_concurrent=...)) — the bridge fork is gone.
• Orchestrator: Env always loads its taskset client-side — info(), run_group(), and task_idx addressing are gone with the protocol. requires_group_scoring and all its consumers go: permit cost is always one episode (acquire()/release() are unit-valued, InflightRollout.rollout_count removed, the run_group fan-out in handle_completed_rollout removed), the group-scored-partial drop policy in the train sink is gone, and the eval sink's buffered-count carve-out simplifies.
• TrainSource / EvalSource: rows always carry the task; next_example() takes no permit argument, and the pending-draw head-of-line hold (which only existed for multi-permit group draws) is removed — pending_env leaves the checkpoint state (old checkpoints load fine; the key is ignored).
• GroupState.task is required; task_idx derives from task.data.idx for error markers.
• Env renames: all taskset ids drop the -v1 suffix across configs/, examples/, k8s/, docs, skills, and the uv workspace members; proposer_solver_v1 import in the algorithm config follows.
• Ports the qwen3-vl features-payload test from the v0 RendererClient to renderers.client.generate — the path the v1 train client actually drives.
• Pins the deps/verifiers and deps/research-environments submodules to the merged companion commits, floors verifiers>=0.3.1.dev14, and relocks. Unit suite passes against the renamed packages (596 tests; the tests/unit/train modules need the flash-attn extra and were skipped locally).

Breaking

• [legacy] config blocks are rejected: any TOML with legacy = { id = ... } under a train/eval source or env-server config fails validation. Migration: port the env to a v1 taskset and set env = { taskset = { id = "<id>" } }.
• Taskset ids drop -v1 in all configs (e.g. reverse-text-v1 → reverse-text).
• Checkpoint note: TrainSource.state_dict() no longer records pending_env; resuming an old checkpoint ignores the key (only ever set for group-scoring v0 envs).

Both companions are merged: deps/verifiers pins main d53ab56f (verifiers#2303 + #2311 + the #2314 catalog revert, released as 0.3.1.dev14 — the floor here) and deps/research-environments pins main c0a0d1d7a (research-environments#754).

🤖 Generated with Claude Code

---

Note

High Risk
Removes legacy env and task id compatibility across orchestration, configs, and checkpoints (pending_env ignored on resume); misconfigured or unmigrated TOMLs will fail validation or resolve wrong tasksets at runtime.

Overview
Breaking change that aligns prime-rl with verifiers v1-only envs and renamed tasksets: configs must use env.taskset.id (no [legacy]), and ids like reverse-text-v1 become reverse-text everywhere (configs, examples, docs, workspace members).

Orchestrator is simplified to a single dispatch path: tasksets load client-side, each rollout is one run with task_data (no task_idx, run_group, or requires_group_scoring). Permit accounting is one episode per in-flight task; train/eval sources no longer gate scheduling on permit cost or hold a pending_env draw.

Config/runtime: EnvConfig / EnvServerConfig drop legacy fields and validators; env-server and RL launcher TOMLs no longer emit legacy. hierarchical_grpo imports proposer_solver instead of proposer_solver_v1. tasks_per_minute is documented as a global dispatch limit.

Deps/tests: verifiers>=0.3.1.dev14, workspace env package paths renamed; Qwen3-VL e2e test uses renderers.client.generate instead of v0 RendererClient.

^{Reviewed by Cursor Bugbot for commit 2583a10. Bugbot is set up for automated code reviews on this repo. Configure here.}