fix(simulation): add_object's is_static is checked, not read by truthiness

[cagataycali]

What

is_static is tri-state, and the resolution that reads it tests by identity (is_static is False / is None) while every later read is a truthiness one. A supplied value outside that domain escaped both, on all three backends. Measured on a real compiled MuJoCo model, one add_object per case:

is_static=	before	body built	after
False	error (documented)	-	unchanged
0 on a plane	success	silently overridden to static	refused
np.False_ on a plane	success	silently overridden to static	refused
"false" on a box	success	welded, no freejoint	refused
None / True	success	as documented	unchanged

0 is the value that refusal answers (int(False) == 0), so one request got two verdicts by spelling — and the accepted spelling got exactly the quiet override the refusal exists to prevent. np.False_ is a value this flag's own domain (boolean_flag_error) accepts, and it took the same path because np.False_ is False is False.

The truthy half inverted the other way. Both cubes below ask for a dynamic body; the left spells it "false", the right False:

Both returned status="success". The "false" body fell 0.0000 m in 400 steps; the False body fell 0.4751 m. 'false' was then stored on SimObject.is_static, annotated bool and read by list_objects, the scene rebuild and domain randomization. Newton and Isaac read the flag purely by truthiness, so "false" fixed a body there too and 0 was stored verbatim.

Why

tests/simulation/test_add_object_static_default_contract.py already pins the plane refusal and names silently overriding it as "the obvious wrong fix... the caller would be told a dynamic plane was built and get a static one". That outcome was reachable through the refusal.

Tests

A supplied is_static now routes through the shared posture domain (SimEngine._validate_posture_flags) on MuJoCo, Newton and Isaac — so a spelling one backend refuses is refused by all of them — and is normalized to a plain bool so the identity reads below it are sound. None stays the "unspecified" sentinel; the plane refusal, its wording and its named remedy are untouched.

69 cells added to the file that owns this contract, table-driven over the falsy/truthy/honored partitions and covering all three backends (GL-free stubs for Newton/Isaac). 35 fail on main, all 66 pass after. The over-refusal cells (True/False/None/np.True_/np.False_ honored, the two postures being genuinely different physics) pass both before and after — that is their job.

Local gate: ruff check + format --check clean on 1,985 files; mypy Success, 0 errors; tests/simulation/ + docstring/refusal graders 16,469 passed, 199 skipped; the 16 other add_object/is_static consumers 1,285 passed.

Round changelog

Round 1 (4dc50cbd + base merge 2bc1c8e5)

• call-test-lint was red on one whole-tree grader, tests/test_no_vacuous_comparisons.py: the new test file asserted 0 == False as a stated premise, and a comparison with a literal on both sides is decided when it is typed, so the grader refuses it. Dropped that one line; the docstring already states the premise and the two add_object calls beneath it are the measurement. This is the PR Workflow step 2 shape (narrow selection green, whole-tree grader red) - noted for the local gate above, which was a tests/simulation/ selection.
• Detect an untested overlap with the base branch was red because fix(simulation): create_world's ground_plane is checked, not read by truthiness #3553 (create_world's ground_plane) landed on mujoco/simulation.py and newton/simulation.py after this branch's merge-base. Merged main into the branch: git merge-tree reported zero conflicts, git show --cc on the merge commit is 0 lines, and the PR's own diff against main is unchanged at 7 files, +296/-3. Composed tree: this PR's two test files 228 passed, fix(simulation): create_world's ground_plane is checked, not read by truthiness #3553's test_create_world_ground_plane_posture_flag_domain.py passed alongside.

[yinsong1986]

Summary

Closes the truthiness gap on add_object's tri-state is_static: the plane refusal tested by identity (is_static is False) while every later read was a truthiness one, so 0, "" and np.False_ reached the silent static override that refusal exists to prevent, and truthy strings ("false", "no", "off") welded bodies asked to be dynamic on all three backends. A supplied value now routes through the shared SimEngine._validate_posture_flags domain on MuJoCo, Newton and Isaac and is normalized to a plain bool so the identity reads below it are sound; None stays the documented "unspecified" sentinel and the plane refusal's wording is untouched.

What's good

• Guard placement is correct on all three backends: before any registry write, prim construction or solver rebuild, so a refusal leaves nothing behind (AGENTS.md > Review Learnings #86 > "Clean up on failure" is satisfied by never registering).
• Follows AGENTS.md > "Posture flags are checked, never read by truthiness" precisely, including the guard-where-the-caller-supplied-it rule and the numpy-bool normalization before the record annotated bool.
• Verified locally at head: all 66 tests in the contract file pass; with strands_robots/simulation/ reverted to the merge base and the head's tests kept, 35 fail — the pins do fail on pre-fix code, matching the PR description exactly.
• Over-refusal guards (True/False/None/np.True_/np.False_ honored, the two postures measured as genuinely different physics) pass on both sides, so the declared domain cannot narrow silently.
• Whole-tree graders spot-checked green at head: docstring xref roles, ASCII tool-result strings, test-name-describes-behaviour; ruff clean on every touched file; changelog fragment named with the PR number.

[yinsong1986]

Summary

Closes the truthiness gap on add_object's tri-state is_static on all three backends. The MuJoCo plane refusal tested the flag by identity (is_static is False) while every later read was a truthiness one, so 0, "" and np.False_ reached the silent static override that refusal exists to prevent, and truthy strings ("false", "no", "off") welded bodies asked to be dynamic — with the raw value then stored on SimObject.is_static (annotated bool). A supplied value now routes through the shared posture domain (SimEngine._validate_posture_flags → boolean_flag_error) on MuJoCo, Newton and Isaac, so a spelling one backend refuses is refused by all, and is normalized to a plain bool so the identity reads beneath it are sound. None remains the documented "unspecified" sentinel and the plane refusal, its wording and its remedy are unchanged.

What's good

• Exactly the AGENTS.md > "Posture flags are checked, never read by truthiness" shape: shared domain, refusal ahead of any side effect (nothing registered on refusal, asserted by the tests), structured error dicts throughout — never raises.
• The guard sits before the plane resolution, so is_static=0 gets the boolean refusal rather than the override, while np.False_ (a value the domain accepts) is normalized and correctly reaches the plane refusal — both directions pinned.
• Over-refusal guard cells (True/False/None/np.True_/np.False_ honored, plus the two-postures-are-different-physics non-vacuity check) pass pre-fix and post-fix, so the narrowing is exactly the non-boolean spellings.
• Cross-backend parity is table-driven and GL-free via the existing _newton_stub/_isaac_stub, extended with the inherited guard so it reads as exercised rather than absent.
• Verified on the head tree: tests/simulation/test_add_object_static_default_contract.py — 66 passed. Changelog fragment follows the <pr-number>-<slug>.md convention; added lines are ASCII-clean with no host paths.