feat: complete per-env fixed model variant roadmap

[TATP-233]

Summary

• Add the task-owned fixed model variant catalog and cold-path Manager-Based lifecycle guard (ADR-0010). The ablated owner API keeps only the variant/catalog config types; Manager factory directly builds UniSim's one final plan.
• Consume UniSim per-world reset defaults in ResetStateTransaction and remove the legacy DomainRandomizationManager/provider protocol; Manager-Based events are now the sole DR lifecycle in this line.
• Add a deterministic representative SimToolReal mesh-variant workload for CPU and MJWarp, including immutable assignment, reset-time mass/inertia DR, per-env playback, cross-backend comparison, and benchmark evidence.
• Ablate the representative integration: expose only materialized model files, reuse the production RSL-RL adapter and shared memory profiler, and avoid duplicating backend conformance assertions.
• Consume the published unisim-core 1.3.0 and mjbatch-uni 0.2.0 registry releases across default and ROCm lock profiles.
• Physically remove the remaining unilab.dr compatibility namespace and provider API docs; Manager-Based construction/reset/interval paths are the sole DR lifecycle, and unsupported capabilities fail closed.

Roadmap: #1563
Representative rollout: #1566
Contract child issues: #1564, #1565, #1567

Validation

• uv sync --locked --extra mujoco --extra mjwarp --extra motrix
• uv run --no-sync pytest -q — 1592 passed, 26 skipped
• make test-all — formatting, mypy, pyright, test checks, 1592 tests with coverage, and benchmark smoke passed
• Representative 64-env / 64-variant benchmark:
  • MuJoCo: median construction 0.181 s over three runs, construction RSS delta about 48.7 MB, all finite
  • MJWarp on RTX 4090: construction 0.970 s, construction RSS delta 420.6 MB, 68,815 env-steps/s, all finite

Integration notes

This PR keeps the package version at 1.2.0 and consumes published registry releases:

• unisim-core>=1.3.0
• mjbatch-uni~=0.2.0

Both default and ROCm locks are generated from PyPI artifacts; no integration-only direct dependency remains.

This PR targets main only, intentionally keeps unilab at version 1.2.0, and does not create a tag, GitHub release, or PyPI publication.

[TATP-233]

Final-head update at 525853735d09144c5e7fb3f1adb8bccf6290dd5b:\n\n- Added the missing end-to-end training-rollout gate: a slow representative 12-env PPO run completes one learning iteration over fixed tool variants and confirms optimizer updates plus finite actor/critic parameters.\n- Current-head CI passed: ruff format/lint, mypy, pyright, benchmark smoke, docs, and Ubuntu coverage test job. Run: https://github.com/Motphys/UniLab/actions/runs/34752848784\n- Targeted local gates also passed for the three representative runtime tests and the slow PPO test.

[TATP-233]

A/B owner/test ablation is complete at a21d6953d551198f116b205d8fa76da54a57e4bd (net -296 lines from the prior head).\n\nUniLab API simplification:\n- fixed-variant task config now exports only FixedModelVariantCfg and FixedModelVariantCatalogCfg;\n- removed FixedModelVariantAssignmentCfg, FixedModelVariantMaterialization, and the public materialization/prepare/validation helpers;\n- empty explicit_variant_names means deterministic round-robin; a non-empty list is the exact expanded assignment;\n- Manager factory directly builds UniSim's plan, so FixedVariantPlan.assignment is the sole final assignment representation;\n- capability negotiation now checks the complete plan (support + layout), not just a single boolean.\n\nRepresentative workload/test simplification:\n- SimToolReal source set exposes only materialized model_files; synthetic variant metadata is private;\n- PPO smoke reuses the production-owned uni_rl RSL-RL adapter instead of adding another test wrapper;\n- benchmark RSS/peak reporting reuses the shared memory profiling helper;\n- representative tests no longer duplicate UniSim immutability/capability conformance assertions;\n- reset-default/entity readonly bindings avoid copying already immutable arrays repeatedly.\n\nValidation:\n- uv run --no-sync pytest -q: 1593 passed, 26 skipped;\n- targeted slow PPO test: 1 passed;\n- make test-all: formatting, mypy, pyright, full coverage suite, and benchmark smoke passed;\n- current PR CI: ruff format/lint, mypy, pyright, benchmark smoke, Ubuntu coverage tests, and Sphinx all passed.\n\nPost-ablation 64-env/64-variant spot benchmark remained finite and functionally complete. CPU median construction over three runs was ~0.181 s with ~48.7 MB construction RSS delta; MJWarp construction was 0.970 s with ~420.6 MB delta. The small timing cost buys shared-field fail-closed validation and MJWarp derived-constant correctness; memory did not regress.

[TATP-233]

Closing cleanup at 6d1fef60271700527b15b6947f6d612735563961: unilab.dr is now physically absent and import fails with ModuleNotFoundError; its API-reference page/card and active provider-language docs are removed. Manager-Based construction/reset/interval paths are the sole DR lifecycle, and the UniSim dependency removes the old init-time model-variant adapter path. Unsupported capabilities continue to fail closed with backend/term-specific errors. Final local gate: 1592 passed / 26 skipped and make test-all passed; current CI and Docs checks are green.

[TATP-233]

Final dependency-only head is 1e772530a3969a92765dc06f2759c1ac94a5ae5d, pinning UniSim a020f8b after its stale DR-manager doc cleanup. Current CI and Docs checks passed again; no main changes.

[TATP-233]

mjbatch-uni 0.2.0 is now published. Head 7564d029389f7fcaff4120018f53b958caea7080 consumes mjbatch-uni~=0.2.0, bumps UniLab to 1.3.0, records the release changelog, and temporarily pins release-candidate UniSim 2060563. Full local suite and make test-all passed; current CI and Docs checks passed.

[TATP-233]

UniSim 1.3.0 is now published. Head eb3e664d58fd9424781775e64f171079f434cdc3 removes the final UniSim git pin and consumes unisim-core>=1.3.0; both default and ROCm locks resolve from PyPI. Full local suite and make test-all passed against the published artifacts.

[TATP-233]

Version bump withdrawn at `c414cbf8d5ce9f3877af76712985a4bdea692c0b`: `pyproject.toml`, `pyproject.rocm.toml`, both locks, and the changelog heading restore/keep UniLab at 1.2.0 / Unreleased. Dependencies remain `unisim-core>=1.3.0` and `mjbatch-uni~=0.2.0` from PyPI. This PR targets main only; no UniLab tag, release, or publication is performed.