docs(roadmap): mark cross-platform infra P0/P1 items as fixed

Five hardening items shipped in this session move to [x] with brief
descriptions + pointers to the regression tests that pin each fix in
place:

  P0 — State Cache Consistency
  P0 — set_states silent-drop guard
  P0 — Robot-config validation tests
  P0 — Backend contract enforcement
  P1 — Parallel-sim error handling
  P1 — Newton joint-name parity
  P1 — Newton silent-action drop guard
  P1 — Actuator partial-spec warning (MuJoCo + Newton)
  P1 — Benchmark reproducibility (env.reset(seed=N))

Each entry has an explicit forward / backward compat statement so
adopters can verify no behaviour surprises. The full migration
write-up lives in MetaSim
``docs/source/developer_guide/architecture_review.md`` v1.1.

Author: geng-haoran

ROADMAP.md

@@ -15,16 +15,65 @@ This document provides a high-level overview of the RoboVerse improvement roadma
 
 ### P0 - Critical (Immediate)
 
-- [ ] **State Cache Consistency** - Fix data corruption when switching state formats
-- [ ] **Test Coverage** - Add validation tests for robot configs and core tasks
+- [x] **State Cache Consistency** - Fixed: dual tensor/dict caches in `BaseSimHandler`,
+  invalidated on `set_states` / `set_dof_targets` / `simulate`. Regression tests:
+  `metasim/test/sim/test_state_modes.py` (integration) + `metasim/test/test_set_states_key_validation.py`
+  (unit, no sim env).
+- [x] **`set_states` silent-drop guard** - Added: unknown / control-input keys now log a
+  one-shot warning (e.g. `dof_pos_target` → "use `set_dof_targets`"). Catches the silent
+  no-op bug that broke 15 downstream BC experiments.
+- [x] **Robot-config validation tests** - Added static `-k general` tests for every shipped
+  `RobotCfg`: instantiation, non-empty `name`, default joint positions match `joint_limits`
+  keys + lie inside the limit ranges. Surfaces 9 real config bugs (xfail-documented:
+  AlohaAgilex/G1Tracking orphan joints, YamCfg/ArxL5Cfg/Vega/SoArm100/Koch/Go2/AllegroHand
+  out-of-range defaults). `metasim/test/test_robot_cfg_validation_general.py` +
+  `tests/test_roboverse_robot_cfg_validation.py`.
+- [x] **Backend contract enforcement** - Static `-k general` test asserts every concrete
+  `BaseSimHandler` subclass overrides each documented contract method. Known incomplete
+  backends (pyrep, partial pybullet/genesis) are xfail-documented. Self-check guards the
+  xfail list against staleness. `metasim/test/test_backend_contract_general.py`.
 - [ ] **CI Coverage Reporting** - Integrate pytest-cov with Codecov
 
 ### P1 - High (Next Release)
 
-- [ ] **Abstract Method Declarations** - Complete @abstractmethod annotations
+- [x] **Parallel-sim error handling** - Fixed: every public method on `ParallelHandler`
+  now drains `error_queue` + checks `process.is_alive` after wire I/O. EOFError/BrokenPipe
+  on recv is translated into a `RuntimeError` carrying the real worker traceback instead
+  of a cryptic IPC exception. OOM-killed workers (queue stays empty) are surfaced via the
+  dead-process check. Tests: `metasim/test/test_parallel_error_handling_general.py`.
+- [x] **Newton joint-name parity** - Fixed: Newton's `_collect_joint_names` now strips
+  the object prefix (`box_base/box_joint` → `box_joint`) when the bare name is unambiguous,
+  matching mujoco / sapien3 / isaacgym / isaacsim. Cross-platform tasks doing
+  `dof_pos["box_joint"]` previously silently broke on Newton. Falls back to the full key
+  on collision so no information is lost. Verified by `test_dict_state_all_objects[newton-*]`
+  flipping from failed to passed.
+- [x] **Newton silent-action drop guard** - Added: Newton's `_set_dof_targets` now warns
+  once if `model.control()` has no `joint_target_pos` / `joint_target_vel` / `joint_f`
+  buffer — previously every action was silently swallowed. Same antipattern that broke
+  MuJoCo `set_states`.
+- [x] **Actuator partial-spec warning (MuJoCo + Newton)** - Identified as task #6 root
+  cause: when an actuator cfg overrides `stiffness` / `damping` but doesn't set
+  `effort_limit_sim`, the asset-authored force-range (MJCF `forcerange` / Newton
+  `joint_effort_limit`) silently dominates. Same `stiffness=1e5` config produces
+  different effective behaviour per backend. Both handlers now warn once per
+  (robot, joint) so the asymmetry is visible.
+- [x] **Benchmark reproducibility — `env.reset(seed=N)` now actually reseeds** -
+  End-to-end wired: `GymEnvAdapter.reset(seed=)` → `RLTaskEnv.reset(seed=)` /
+  `TaskBase.reset(seed=)` → `BaseSimHandler.set_seed`. Added a default
+  `set_seed` on the base handler that reseeds Python `random` + NumPy +
+  Torch (CPU and CUDA), so every backend gets reproducibility for free.
+  Backends with extra internal RNG (Newton warp kernels, Sapien physics
+  noise) can override and call `super().set_seed(seed)` first. Tests:
+  `test_set_seed_is_deterministic_on_numpy_and_torch` and
+  `test_set_seed_differs_with_different_seeds` in
+  `metasim/test/test_set_states_key_validation.py`. Backward-compat:
+  `seed` is a keyword-only parameter — every existing call site is
+  unchanged.
+- [ ] **Abstract Method Declarations** - Mostly covered by the new backend-contract test;
+  still want `@abstractmethod` on `_get_joint_names` / `_get_body_names` once pyrep /
+  partial pybullet / genesis catch up.
 - [ ] **Unified Environment Factory** - Create `roboverse.make_env()` API
 - [ ] **Configuration System** - Document and standardize config approaches
-- [ ] **Error Handling** - Improve parallel simulation error propagation
 
 ### P2 - Medium (Future)