chore: prune redundant tests and remove legacy task-factory seam

[TATP-233]

Summary

• Removed redundant tests identified by a full tests/ audit (7 parallel read-only audits, every candidate re-verified against src/ + git history before removal): old-interface guards (reward_config/domain_rand/control_config legacy-override loops, removed offpolicy option/kwarg guards, DrakeUniBatchBackend/A2JoystickCfg absence assertions, _selected_sim_backend/obs_profile markers), strict-duplicate coverage (g1 owner compose tests, motrix-sac/td3 compose tests, reward-injection integration, migration-matrix subset, APPO init-no-crash, observation finite-policy subset), tautological tests (former-script frozen-copy parity family, mock self-test, no-signal compile override, self-referential obs-alignment file), tests of third-party internals with zero unilab contact (unisim mjwarp private suites, body-state copy, MjSpec compile), and dead conftest fixtures (mp_ctx/tiny_storage/tiny_weight_shapes).
• Removed the legacy task-factory seam src/unilab/tasks/compatibility.py (zero production users): deleted its test file and the closeout seam test, dropped the "compatibility" migration target, and rewrote the registry closeout so it enforces that legacy EnvCfg -> NpEnv factories cannot coexist with the canonical Manager-Based runtime.
• Replaced the outdated private _RslRlVecEnvWrapper copy in test_rsl_rl_runner.py with the production uni_rl.algos.rsl_rl.RslRlVecEnvWrapper.
• Updated 6 docs pages (en + zh_CN) that cited the deleted obs-alignment test to reference the real guarantee (ObservationManager per-term history buffers, tests/managers/test_observation_buffers_noise.py).
• Unique assertions from deleted duplicates were folded into the superset suite (g1 owner contract now pins fragment_files/terrain/genesis_device_id/pd_gains is None for isaacgym/genesis owners).
• No user-facing, training, backend, or runtime behavior change. 41 files, +44/−1867.

Linked Work

• Issue: chore: prune redundant defensive tests and remove the legacy task-factory seam #1572
• Parent roadmap (when applicable): none
• Roadmap declared base (when applicable): n/a
• Milestone: n/a
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created or updated — partially blocked by a pre-existing local environment issue, see below
• Additional task-specific validation listed below

Commands actually run (on final head 98df4068):

make check                  # PASS: ruff format/check, mypy 137 files, pyright 0 errors, tests lint
make test-all               # check PASS; test-cov ABORTED at collection: 7 files error with
                            # AttributeError: module 'mjbatch' has no attribute 'VariantPack'
make test-benchmark-smoke   # PASS: 34/35 (1 skipped, platform-optional mlx)

# Failure-set equivalence vs main (same command on both heads, sorted FAILED lists diffed):
uv run pytest -m "not slow" -q --continue-on-collection-errors \
  --ignore=tests/algos/test_appo_runner.py --ignore=tests/algos/test_rsl_rl_runner.py \
  --ignore=tests/base/backend/test_mujoco_scene_context_visual.py \
  --ignore=tests/base/test_sim_backend_set_state_timing.py \
  --ignore=tests/base/test_sim_backend_smoke.py \
  --ignore=tests/scripts/test_train_script_configs.py \
  --ignore=tests/training/test_training_helpers.py
# main: 40 failed / this head: 39 failed — diff shows the ONLY delta is the deleted
# tests/base/test_dr_legacy_removed.py (itself failing on main). Zero new failures.

uv run pytest tests/scripts/test_check_docs.py -q          # PASS: 19
cd docs/sphinx && UNILAB_DOCS_SKIP_AUTODOC=1 uv run --no-project \
  --with-requirements requirements.txt sphinx-build -b html -n source build/html   # PASS

The test-cov collection errors are a pre-existing local environment defect, not caused by this PR: the venv has mjbatch-uni 0.1.0 while pyproject.toml requires ~=0.2.0, so every file that imports unisim.backend.mujoco.backend fails identically on main (verified via git stash baseline: same 7 collection errors, same 39+1 runtime failures). A clean CI environment is the real gate for those files.

Remote CI route:

• Base main: current-head CI running, see checks below (link: https://github.com/Motphys/UniLab/pull/1573/checks).

Impact

• Backend impact: none
• Platform impact: Linux (validated); no platform-specific behavior touched
• Training effect expected: no

Artifacts

• W&B: n/a
• benchmark result: make test-benchmark-smoke 34/35 (1 platform-optional skip)
• video / screenshot: n/a
• ONNX / checkpoint: n/a

Checklist

• Added or updated tests where needed (registry closeout rewritten; g1 owner contract strengthened)
• Updated docs if behavior or workflow changed (6 pages, en + zh_CN)
• Linked the driving issue (chore: prune redundant defensive tests and remove the legacy task-factory seam #1572)
• Noted any follow-up work explicitly:
  • Local env needs mjbatch-uni ~=0.2.0 (reinstall/rebuild) for the mujoco-backend test files; affects main equally.
  • Kept deliberately: skip-gated drake_uni acceptance tests (zero CI cost, only acceptance signal when the optional extension is installed) and the registry's legacy reward_config routing tests (live code for external plugin packages).