Rename the predicatorv3 configs to empiric with rounds and subsets

[yichao-liang]

Summary

scripts/configs/predicatorv3/ (51 files) becomes scripts/configs/empiric/, which holds only the paper's experiment:

• common.yaml: the shared flags, including the principled joint belief (belief_joint_draws: 16, no carried posterior).
• envs.yaml: the five benchmark settings (Balloons composition, four-span Bridge, two-jug Boil, ramp Fan, Domino high-friction turn).
• approaches.yaml: the seven arms (EMPIRIC, model-free, model-free + scene package, standalone, oracle dynamics, no fitting, no uncertainty).
• benchmark.yaml: the default launcher, which includes the three, over seeds 0-4.

With --round joint_r1, benchmark.yaml resolves to the same 35 runs as continual_principled_belief_r1.yaml did: same experiment ids, approaches, envs, args and flags. The only differences are seeds 0-4 instead of 0-1, and two top-level flags that equal their defaults.

Launching:

• A launch names its round with --round or a ROUND key, which suffixes every experiment id (<env>-<arm>_<round>). launch.py refuses a continual launch without one, since those runs auto-resume from their run folders.
• --envs, --approaches and --seeds N-M pick a subset.
• A launch worth keeping on record can still get its own YAML that includes benchmark.yaml and sets ROUND.
• EXTENDS, which gave arms round-specific ids, is removed.

Deleted:

• the phased exp_*.yaml configs, their all.yaml menus, common.yaml and oracle.yaml;
• the dated continual_* launchers and the menu entries outside the benchmark (the scene-only, zero-shot, from-assets, real-to-sim, EMPIRIC + scene package and Sonnet entries; their classes stay);
• scripts/engaging/relaunch_on_timeout.py, whose job the self-requeue trap now does;
• scripts/domino_debug/.

random_actions_pybullet.yaml moves next to the ExoPredicator configs it belongs with, and generate_random_action_gifs.py now finds it.

Tests that loaded dated launchers now load the benchmark. The arms outside it keep their tests, with their flags defined in the tests. Doc links to deleted launchers point at the files under the iclr-empiric-submission tag, and code comments that pointed at the band probe name it at that tag too.

Test plan

• yapf, isort 5.10.1, docformatter, mypy and pylint on this commit (sbatch replay of CI's static jobs).
• The 8 CI shards in the CI container. The first run caught a recursion in the comparison test's new config helper; after the fix, test_continual_comparison_approach.py passes (41 tests) and the other shards' failures were only that helper.
• benchmark.yaml --round joint_r1 against continual_principled_belief_r1.yaml: identical runs apart from the seeds.

🤖 Generated with Claude Code