domino_real: fold perceived roll modulo pi

[amburger66]

A bug found while running the oracle approach on the real-scene domino env against a real capture.

Symptom

Every domino on the table was standing, but the run reported success without moving:

Task init atoms: [... Toppled(domino_0:domino), Toppled(domino_2:domino)]
Found Plan of length 0 in 0.00s
Task 1 / 1: SOLVED

The goal {Toppled(domino_0)} was already true in the task's own initial state, so the planner returned an empty plan and the run scored 1.0. A vacuous pass.

The state dump showed two of four dominoes at roll = 3.14159 and the other two at roll = 0. Not noise: exactly π.

Cause

A domino is a box, so turning it 180° about its own width axis leaves it exactly where it was. Both orientations describe the same physical domino, and a marker-based pose estimate returns either one arbitrarily — within a single capture, some dominoes come back at roll 0 and others at roll ±π.

_Toppled_holds is |roll| >= 10°, so an unfolded roll = π reads as knocked over.

This is a regression from the toppled-domino support in plan PR 6 (#95). Before it, state_from_observation and _init_state_from_perceived hard-coded roll = 0.0, so a flipped estimate could not surface. My offline check against scenes 0000/0001 missed it because every domino in those two captures happened to come back near roll 0 — the flip is capture-dependent.

Fix

Fold roll into [-π/2, π/2):

• standing (0 or ±π) → ≈ 0 → Upright
• knocked over (±π/2) → magnitude preserved, which is all Toppled/Upright read

Yaw is untouched, so the push direction is unaffected. The fold lives in the domino env rather than in domino_env_euler, so the geometry helper stays a faithful decomposition and the "a domino is a box" fact sits with the domain that knows it.

Verification

Against the capture that exposed this — all four dominoes now read 0.00°, Upright, and the goal is no longer satisfied at init:

obj        capture id  roll(deg)   Upright  Toppled
domino_0            5       0.00      True    False
domino_1            1       0.00      True    False
domino_2            2       0.00      True    False
domino_3            6       0.00      True    False
goal already true at init: False

Note the earlier SOLVED should not be read as the scene being solvable — that question is still open, and re-running is the way to answer it.

Tests

Three additions to tests/envs/test_pybullet_domino_real.py:

• the invariant: a capture whose records are 180°-flipped still builds a task with every domino upright, and the goal not already satisfied by its own initial state;
• the same for a mid-episode observation, so the fold applies to state_from_observation too;
• the fold itself over eight angles, including that ±90° keeps its magnitude.

Removing the fold fails 6 tests, including both regression tests.

Gates: 39 in this file (was 29), tests/envs + tests/pybullet_helpers = 297 passed / 15 skipped; mypy clean over 709 files; pylint 88/88; isort/yapf/docformatter verified by exit code.

[yichao-liang]

LGTM!