Reject wrongly shaped params in ParameterizedOption.ground

[yichao-liang]

ParameterizedOption.ground now raises a clear ValueError when the params' shape differs from the params space's shape, for example Cannot ground 'PickJug': params [] have shape (0,), expected shape (1,).

Why. When params_space.contains(params) failed, ground fell back to a float-tolerance bounds check (added in #57) that never looked at the shape.
Empty params for a one-parameter space passed it vacuously (np.all over no values is True), so the option was grounded with no params and failed later.
This is how Grow's oracle crash got through grounding before #201 fixed its samplers: the NSRT's null_sampler gave the skill-factory PickJug no params, and pick.py's _descend_pose raised an IndexError reading params[0].
For a space with more than one entry, the bounds comparison raised a bare broadcast error instead.

The sampler path had the same hole. _GroundNSRT.sample_option and _GroundEndogenousProcess.sample_option clip sampled params with np.clip(params, low, high) before grounding, and np.clip broadcasts.
It turned a (1,) sample into (0,) for a parameterless option and copied one value into all four params of a (4,) space.
For empty params and a (4,) space it raised the operands could not be broadcast together with shapes (0,) (4,) (4,) error seen with Coffee's PlaceJugInMachine, before ground was even called.
Both now clip only correctly shaped params (_clip_sampled_params) and pass anything else to ground, which rejects it.
default_params substitution for empty params is unchanged, so Wait options still ground from [].

Callers that relied on the old behavior, found by the full suite and by a sampler-shape audit over every ground-truth NSRT and endogenous process (default config and each env's legacy options), now fixed:

• utils.ops_and_specs_to_dummy_nsrts gave every NSRT the dummy sampler np.zeros(1) and relied on the clip to reshape it; test_llm_open_loop_approach grounded its parameterless PDDL options through it.
  Its samplers now return zeros of each option's params shape, which the clip turns into the same values as before.
• Grow, Boil and Coffee pick processes returned a grasp offset [x] even with legacy options, whose PickJug takes no params; the clip used to drop it silently.
  They now return [] with legacy options, as their place and push samplers already do.
• Coffee's TwistJug process used null_sampler although TwistJug takes a twist amount when coffee_twist_sampler is on (the default); it passed only through the vacuous check.
  It now draws the amount as the TwistJug NSRT's sampler does.

Left as is. On top of #201, which fixed the Grow and Coffee NSRT samplers, the audit finds only legacy Grow's Place (with grow_use_skill_factories off) still sampling the wrong shape.
Its PlaceJugOnTable process gives [] to the two-parameter Place, and with grow_place_option_no_sampler on, its PlaceJug NSRT gives [x, y] to the parameterless one.
Both already raised broadcast errors on master and now raise the clear error; legacy Grow does not plan on master in either mode.
LinearChainParameterizedOption's consistency check has the same hole (np.allclose on (0,) and (1,) bounds passes), and Coffee's combined Twist chain relies on it, so it is left for a separate change.

Test isolation fix. pytest tests/test_structs.py on its own errored in 5 tests, and tests/agent_sdk/test_bilevel_sketch_samplers.py on its own failed test_refine_and_validate_report_returns_plan.
State.pretty_str() reads CFG.excluded_objects_in_state_str, an args.py setting that exists only after a reset_config() call, so these tests passed only when an earlier test in the process had called it.
A session-scoped autouse fixture in tests/conftest.py now calls reset_config() once before the first test.

Test plan

• New test_option_ground_rejects_wrong_params_shape (the Box(0, 0.1, (1,)) option grounded with np.array([]), plus (4,), (1, 1) and scalar cases, parameterless options and default_params) failed on master and passes here.
• New test_sample_option_rejects_wrong_params_shape covers NSRTs and endogenous processes: wrong shapes raise, correct shapes are still clipped, empty params still take default_params.
• Full suite on the final commit (rebased on master 423cae0c1), split round-robin by test file into 16 jobs in CI's Ubuntu 24.04 container: 2546 passed, 0 failed (4030 skipped, 14 xfailed, 4 xpassed).
  The first run, with only the ground check, found the order-dependent test_bilevel_sketch_samplers failure (also on master); the second found test_llm_open_loop_approach (the dummy samplers).
• New test_process_samplers_match_option_params (Grow, Boil and Coffee, skill-factory and legacy options) fails on master in exactly the four fixed cases and passes here.
• Sampler-shape audit over every ground-truth NSRT and process (default and legacy configs, plus Coffee's twist variants): the only mismatches left are legacy Grow's, listed above.
• The README command (main.py --approach oracle and oracle_process_planning, seed 0, one task, 60 s) on Grow and Coffee ends as on master 423cae0c1: Grow's oracle times out in refinement and its process planning solves 1/1; Coffee fails on cup0's pose with both.
• Before Sample the skill-factory parameters in Grow's and Coffee's NSRTs #201 (at 864ed48c7), Grow's oracle raised IndexError in _descend_pose; with this change it stopped at grounding with Cannot ground 'PickJug': params [] have shape (0,), expected shape (1,).
• yapf 0.32.0, isort 5.10.1, docformatter 1.4, mypy 1.8.0 and pylint on a clean clone of the final commit: no diffs, no issues.

🤖 Generated with Claude Code