fix(eval): enforce official LIBERO and LIBERO-Plus protocols

[primorLee]

Problem

The existing client launcher always uses --task-range 0 10 --test-num 50. That is the original LIBERO protocol (10 tasks x 50 trials), but under LIBERO-Plus those indices are perturbation variants rather than the ten original tasks. The command therefore repeats only the first ten Plus variants 50 times each and can make a 500-rollout subset look like a full LIBERO-Plus score.

This also explains why a valid original-LIBERO checkpoint reproduction and the reported Plus result can diverge without establishing a checkpoint regression: they were not evaluating the same protocol.

What changed

• make the protocol explicit with --protocol original|plus
• preserve original LIBERO behavior and compatibility:
  • 10 tasks x 50 trials by default
  • existing positional run(...) arguments
  • existing per-task result and video paths
• enforce the official LIBERO-Plus protocol:
  • every perturbation variant exactly once by default
  • exact official suite sizes (2402/2518/2591/2519)
  • exact alignment with the pinned official task_classification.json
  • BDDL :language instruction from the environment instead of metadata-bearing variant filenames
  • category/difficulty numerators and denominators
• distinguish diagnostics from reportable scores:
  • completed shards have an observed rate, but score.value remains null
  • interrupted full runs remain non-reportable until every planned variant completes
• add reproducible result lifecycle controls:
  • run fingerprint over protocol, selection, declared checkpoint, client/protocol code, benchmark order/metadata, and package versions
  • immutable manifest and atomic incremental task/summary writes
  • validated resume that rejects corrupt or mismatched same-run results
  • collision-free Plus run directories
  • required user-declared --checkpoint-id NAME@REVISION for Plus, since the inference server is a separate process
• add a dedicated launch_client_plus.sh; both launchers forward extra CLI arguments
• document the two protocols and non-comparable denominators

Validation

• 24 passed protocol/lifecycle tests
• Ruff check and format pass
• Python compilation and diff whitespace checks pass
• live metadata check against LIBERO-Plus commit 4976dc30028e805ff8094b55501d532c48fec182

I did not run a real MuJoCo/checkpoint rollout in this change, so this PR makes no model-success-rate claim. It fixes the evaluation contract needed before rerunning the checkpoint.

References:

• Official LIBERO-Plus evaluation instructions
• LIBERO, NeurIPS 2023 Datasets and Benchmarks

Fixes #115