fix(simulation): replay_episode resolves robot_name the way its siblings do

[cagataycali]

What

replay_episode resolves robot_name through the shared SimEngine._resolve_single_robot, the rule run_policy, eval_policy and evaluate_benchmark already apply: None picks the sole loaded robot, an ambiguous scene is refused listing the candidates, and a supplied name is never re-resolved. The private copy in PolicyRunner.replay (robot_name or <first robot>) goes, with its two unit tests.

Why

The copy read the selector by truthiness and picked a default where the siblings refuse. Two consequences, both reported as success:

call	scene	before	after
replay_episode(ds, robot_name="")	one robot	substituted the only robot and replayed onto it	Robot '' not found in sim. Available robots: [...]
replay_episode(ds, robot_name="")	two robots	substituted the first robot and replayed onto it	same refusal, by name
replay_episode(ds)	two robots	replayed onto whichever robot list_robots() put first, naming it in the success text	Multiple robots registered; specify robot_name. Available: ['arm_a', 'arm_b']
replay_episode(ds)	one robot	resolved the sole robot	unchanged

Replay is the one policy surface that drives the actuators from a recording rather than a policy, so the substitution the siblings refuse is a robot the caller never chose being moved. Measured on the fake engine: two frames of recorded actions reached arm_a in an arm_a/arm_b scene with no name given. #3610 closed the same class on evaluate_benchmark; this is the fourth and last surface in SimEngine that takes a robot_name and drives a rollout.

"" now reaches the existing membership check unchanged, so it is reported by name with the loaded set, the same text an unknown name gets. The replay docstrings said "defaults to the first robot in the sim"; they and docs/recording.md now state the shared rule.

Tests

tests/simulation/test_simengine_facade_guardrails.py:

• test_a_supplied_robot_name_the_scene_lacks_is_reported_by_name gains replay_episode as its fourth row in both scene shapes. The dataset loader is stubbed to fail loudly, so the cell also grades that the refusal precedes the download. Both new rows fail pre-fix (the loader was reached); the six sibling rows are controls.
• test_replay_episode_ambiguous_multi_robot_refuses_before_commanding_anything: the loader is left reachable and send_action is recorded, so pre-fix this fails on the commanded list (['arm_a', 'arm_a']) rather than on a stubbed import.
• test_replay_episode_resolves_the_sole_robot_when_name_omitted: control, two frames reach solo.

gate	result
new cells, pre-fix policy_runner.py	3 failed / 7 passed
new cells, this head	10 passed
the six touched test files + describe suites	302 passed; the 4 failures are mujoco/pyarrow absent locally and reproduce identically on main
ruff check + ruff format --check	clean, 2018 files
mypy on the five touched Python files	clean
whole-tree graders (143 of the roster's 145; the 2 left out import lerobot, absent locally)	4623 passed, 70 failed - failure set byte-identical to main on the same 13 files, 0 attributable

Overlap: #3343 also edits simulation/policy_runner.py, in _resolve_success_fn (~line 4222); this change is in replay (~line 3088) and deletes a helper at ~4581, so it is a merge-order question with no shared hunk.

[yinsong1986]

Summary

Removes PolicyRunner.replay's private robot-resolution copy (robot_name or <first robot>) and routes replay_episode through the shared SimEngine._resolve_single_robot, so the fourth and last robot_name-taking rollout surface now applies the same rule as run_policy / eval_policy / evaluate_benchmark: None resolves the sole robot, an ambiguous multi-robot scene is refused with the candidate list, and a supplied name (empty string included) is never re-resolved and hits the existing membership check. This replaces two silent-substitution behaviours that both reported success with explicit structured errors, which is the correct direction under AGENTS.md's "a subset selector is read by membership, never by truthiness" convention, and closes the class #3610 opened on evaluate_benchmark.

What's good

• Verified locally at c0b97c8d: the three new test cells fail on the pre-fix policy_runner.py (exactly the 3 claimed, 7 controls green) and the touched suites pass at head (114 passed).
• The ambiguous-scene test grades the actual hazard - it records send_action targets and asserts nothing was commanded - rather than only the error text, and the unknown-name rows stub the dataset loader to fail loudly so the refusal is proven to precede the download.
• The deleted helper's two unit tests go with it, test_policy_runner_paths.py's no-robots test still passes against the shared resolver's message, and no remaining caller of _require_default_robot exists in the head tree.
• Docstrings (base.py, policy_runner.py) and docs/recording.md updated in the same change; the new :meth: cross-reference roles resolve (xref grader green at head); changelog fragment named for the PR number is present.