fix(tasks): honor reset(states=) contract in eight Tier-1 task overrides

[geng-haoran]

BaseTaskEnv.reset(states=None, env_ids=None, seed=None) is the contract every task must be
substitutable for, but eight Tier-1 overrides dropped states: the two legged-robot bases
(humanoid, beyondmimic) reimplemented reset without it, and the six SimplerEnv tasks took
options in its place.

The IL and VLA evaluation runners pass states unconditionally — default_runner calls
env.reset(states=init_states[...]) to replay a demo's initial states, and the OpenVLA, SmolVLA
and pi0 eval loops do the same. Because --task is a free-form string, nothing stopped a user
from pairing a VLA with a SimplerEnv task — the canonical pairing for that benchmark — and doing
so died immediately with TypeError: reset() got an unexpected keyword argument 'states'. The
same TypeError hit any IL evaluation on the humanoid or beyondmimic bases.

All eight now accept states/env_ids/seed in the base's order. The legged-robot bases apply
caller-supplied states in place of their scripted initial states. The SimplerEnv tasks keep
options — it carries the episode spec (object/robot init options, episode id) that the gym
adapter feeds them, so dropping it would have silently discarded per-episode setup — but it
moves after the contract's three parameters and keeps its default. External states are applied
on top of the scripted layout, and the settle-derived quantities the success checkers key off
(object height, object poses, bbox extents) are re-derived from the applied states rather than
left describing a layout that was overwritten.

The guardrail that should have caught this (tests/test_task_reset_seed_contract.py) only looked
for a seed parameter, so all eight sailed through it — that is why they rotted. It now checks
full substitutability: every contract parameter is accepted, positional order matches the base
(a swapped positional binds the wrong value, which corrupts rather than raises), and an override
may not add a required parameter of its own. Extra parameters like SimplerEnv's options remain
allowed after the contract's three, with defaults. The ratchet is zero-tolerance, so this class
of drift cannot recur.

Review: independently reviewed against current main (verdict MERGE AS-IS); rebased, re-verified: 14 passed in 6.72s

🤖 Generated with Claude Code

https://claude.ai/code/session_017i6VtKoovBNed815mWFqxw

[cloudflare-workers-and-pages]

Deploying roboverse-release with    Cloudflare Pages

Latest commit:	09720eb
Status:	⚡️  Build in progress...

View logs