feat: add manager-based mjwarp training infrastructure

[TATP-233]

Summary

• 交付 contract-first 的 manager-based task API：冷路径 TaskCompiler 生成不可变 compiled plan，reference、host-fused 与 device-resident executor 共用 typed state/control/mutation contract。
• 新增与 mujoco 完全独立的生产 mjwarp backend identity、factory、owner YAML、registry、telemetry、typed batch/reset/DR、graph recapture 与 device substep controller。
• 打通 ppo_torch + g1_walk_flat + mjwarp + device_resident 从 policy action、physics、obs/reward/done 到 learner 的 device-resident 训练路径，并建立 policy ABI、sim2sim、checkpoint/resume/export 入口门禁。
• 建立 Phase 0-7 claim -> evidence -> gate 体系，覆盖基线、正确性、差分、性能、DR、训练行为、入口矩阵、legacy retirement、support promotion 与最终 freshness。
• 淘汰 G1WalkFlat + mjwarp 的手写 env lifecycle；mujoco / motrix owner 及 reference/fused/differential oracle 保留。

Linked Work

• Issue: Closes Work: 统一规划 manager-based task API 与独立 mjwarp 后端 #705
• Related: Work: 设计 contract-first manager-based task API（参考 Isaac Lab / mjlab） #702, Work: 新增独立 mjwarp 物理后端 #704
• Milestone: Work: 统一规划 manager-based task API 与独立 mjwarp 后端 #705 Phase 0-7 全部 verified
• 中间工作按 Phase 0-7 拆为 child issue 和小 PR，全部以 feat/issue-705-manager-mjwarp 为 base 合入；完整 child issue、依赖和逐 PR 证据见 Work: 统一规划 manager-based task API 与独立 mjwarp 后端 #705 正文。最终阶段对应 Phase 7A: 建立 mjwarp 支持声明与双向证据审计 #833/test: add Issue 705 support evidence audit #834、Phase 7B: 建立 capability-derived mjwarp production task rollout #835/test: add capability-derived mjwarp task rollout gate #836、Phase 7C: 冻结并验证 mjwarp 逐 seed 训练行为 #837/test: verify Phase 7C paired-seed training behavior #838、Phase 7D: 建立 manager/mjwarp 全入口能力矩阵与 ABI 门禁 #839/feat: enforce issue 705 entrypoint contracts #840、Phase 7E: 淘汰 mjwarp 手写 env lifecycle 并冻结回滚证据 #841/feat: retire legacy mjwarp G1 lifecycle #842、Sync: #705 最终回归前同步最新 main #844/chore: sync main before issue 705 final regression #845、Phase 7F: 全阶段 freshness 最终回归与 Phase 7 promotion #843/test: close issue 705 final evidence #846，以及 CI 修复 Fix: 保持 mjwarp armature recompute 的可选依赖边界 #848/fix: preserve mjwarp optional import isolation #849、Fix: 拆分标准 CI 并隔离本地 evidence 门禁 #855/ci: shard standard tests and preserve local evidence #863/fix: make issue 705 CI evidence portable #864。

Support Boundary

唯一 Recommended 组合为：

ppo_torch + g1_walk_flat + mjwarp + device_resident

• mjwarp 与 mujoco 始终是两个独立 physics backend，不共享或改写 backend identity。
• PPO MLX、其他 task 和其他 execution profile 未被提升为 Recommended。
• mjwarp 原生 play/visualization 当前显式 unsupported，并在 materialization 前 fail closed；checkpoint load/resume 和 ONNX/JIT export 走已验证的 renderer-independent contract。
• 未声明的 DR/Event/controller/model field 默认 fail closed。已声明能力由 field-level manifest、selected-row isolation、physics-effect、recompute 和 graph tests 约束。
• RSS 按已批准修订仅作为 post-warmup 稳态诊断，不设置绝对 peak ceiling；transfer、allocation、地址稳定、graph、吞吐与训练行为仍为硬门禁。

Validation

• make check（由最终 make test-all 覆盖；Ruff、mypy、pyright 均 PASS）
• uv run pytest -m "not slow"（由最终 make test-all 覆盖）
• Additional task-specific validation listed below

Commands run against child HEAD b98ef976acc4894e08a9c95634a0a3eb41eb035f and promoted to integration HEAD 480becc726d8c73e06669f4dc02baab2568be742（merge commit only adds ancestry）：

env LD_LIBRARY_PATH=/tmp/nvidia-570.133.07/usr/lib/x86_64-linux-gnu make test-all
uv run ruff format --check .
uv run ruff check .
uv run mypy src/unilab
uv run pyright
uv run scripts/audit_ci_test_shards.py --json
uv run scripts/audit_issue705_workflow_triggers.py --json
uv run scripts/audit_issue705_claims.py --all --json
uv run scripts/audit_issue705_backend_isolation.py --json
uv run scripts/audit_issue705_support.py --json
uv run scripts/audit_issue705_task_rollout.py --json
uv run scripts/audit_issue705_legacy_retirement.py --json
uv run scripts/validate_issue705_dr_inventory.py --json
uv run scripts/validate_issue705_phase.py --phase 4 --mode gate --json
uv run scripts/validate_issue705_phase.py --phase 5 --mode gate --json
uv run scripts/validate_issue705_phase.py --phase 6 --mode gate --json
uv run scripts/validate_issue705_phase.py --phase 7 --mode gate --json
uv run scripts/validate_issue705_final.py --json
uv run scripts/validate_issue705_final.py --head-only --allow-unpromoted --json
git diff --check
git status --short --branch

Results:

• Full gate: 2348 passed, 26 skipped, 449 deselected, 1 xfailed; coverage 68%; Ruff, mypy and pyright PASS.
• Standard CI equivalent shards: A 870 passed, 15 skipped, 281 deselected, 1 xfailed; B 551 passed, 1 skipped, 114 deselected; C 285 passed, 8 skipped, 12 deselected; D 636 passed, 2 skipped, 48 deselected; all PASS.
• CI shard audit: 8 shards / 227 test files / exactly 6 local evidence nodes / 0 errors; final remote merged coverage 67%, --fail-under=25 PASS.
• Benchmark smoke: module mode 36/37, script mode 37/38; only the platform-optional MLX entry skipped in each mode.
• Final gate: ok=true, 12 mandatory command runs, A/B/C/D lanes complete, errors=[].
• Phase 7: 6 verified / 0 planned; Phase 0-6 owner validators PASS.
• Support: 1 combination / 1 Recommended / 0 errors.
• Task rollout: 1 entry / 27 prerequisites / 0 errors.
• Claims: 51 claims / 115 entries / 115 existing / 0 targets / 49 supporting.
• Legacy retirement, backend isolation, workflow trigger, DR inventory and sim2sim audits PASS.
• Child worktree clean; remote integration HEAD is 480becc726d8c73e06669f4dc02baab2568be742; no main-only commit remains after chore: sync main before issue 705 final regression #845.

Remote CI Status

• Real run 30698393464 on 30d69111ebab8dab2743ea4d4299c3ff96c87fa3: Ruff、mypy、pyright、Docs、shard A/B passed; shard C/D and aggregate failed.
• Root causes were shallow Git history for evidence ancestry and Python 3.11/3.13 throughput ULP roundoff. PR fix: make issue 705 CI evidence portable #864 fixed both without changing evidence thresholds or the six-node local allowlist.
• Final CI run 30742715153 and Docs run 30742715151 on integration HEAD 5a8a26691e33bb3c9af1d088c89bb7a36c221d62 both succeeded; all eight shards and aggregate test (ubuntu-slim) are green.

CI Evidence Boundary

• Hardware fields remain required and structurally validated, but CPU/GPU model、UUID、capacity、driver 和 affinity 与冻结 profile 的差异仅作为 provenance warning，不再是 equality gate。
• Artifact、execution、worker 与 profiler 必须对实际记录的硬件和 affinity 保持内部一致；线程环境、source freshness、raw recomputation、process isolation、依赖和矩阵完整性仍 fail closed。
• Phase 5/6/7 的数值性能阈值保持硬门禁；本次未修改 threshold manifest，也未把失败诊断提升为 evidence。

Issue #877 Update

• Fixes Work: 证据链硬件 profile 从 equality gate 降级为 provenance metadata（去除固定硬件 / NVIDIA driver 版本耦合） #877
• 将 Phase 0/4/5/6/7 的冻结硬件逐字段 equality gate 降级为 provenance metadata + UserWarning；当前主机的实际 CPU/GPU/driver/affinity 会写入 artifact。
• GPU memory ratio 使用实际记录的 device capacity；runtime affinity 优先使用本机可用的冻结 CPU 子集，否则记录当前 allowed cpuset。
• correctness、backend、thread environment、freshness、raw recomputation、process isolation 和性能阈值 contract 均未放宽。
• Phase 3 evidence 已刷新；Phase 4 的 45 个 host benchmark 进程、raw artifact 和 gate 已刷新。

#877 Validation

• Targeted suite：196 passed；Ruff format/check、mypy、pyright 和 benchmark smoke 通过（合并最新 integration 前）。
• 合并最新 integration 后：Ruff + mypy 通过；managed policy ABI / RSL-RL adapter 聚焦测试 17 passed, 13 deselected。
• Phase 3 acceptance：2 passed。
• Phase 4：45/45 benchmark workers PASS；acceptance 2 passed。
• Phase 5：40/40 GPU workers PASS，硬件差异仅产生 advisory；冻结性能 gate 失败：
  • batch 128 p50 ratio 1.1596，要求 <= 1.05
  • batch 128 p95 ratio 1.1539，要求 <= 1.05
  • batch 128 FPS ratio 0.8623，要求 >= 0.95238
• Diagnostic artifact 的 non-threshold validation errors 为 ()；未修改或新增 threshold amendment。

Explicit PR Gate Override

用户于 2026-08-03 明确要求“跳过性能阈值的 gate，直接 PR”。因此本次更新在以下状态下推送：

• 未通过最终 make test-all / Issue Work: 统一规划 manager-based task API 与独立 mjwarp 后端 #705 final evidence gate。
• Phase 5/6、support、legacy retirement 和 final evidence 仍有 freshness failure。
• 未把 diagnostic artifact 提升为正式 evidence，未伪造 PASS，也未降低性能阈值。
• Reviewer/merger 需显式接受该 override；后续若要恢复全绿 evidence，需要独立 threshold amendment 或在满足现有阈值的证据主机重新采集。

Impact

• Backend impact: new independent mjwarp; additive shared contracts for mujoco; motrix/drake remain fail-closed where unsupported.
• Platform impact: Linux/CUDA for the Recommended path; optional Warp imports remain cold-path, while standard non-CUDA CI uses static artifact validation.
• Training effect expected: yes. The production G1 PPO route uses manager-owned device buffers and avoids per-step CPU round trips.
• Compatibility: existing NpEnv NumPy contract and non-mjwarp G1 owners remain available; backend-specific behavior stays in backend/config owner layers.

Artifacts

• Final gate: tests/acceptance/issue_705/artifacts/phase_7_gate.json
• Final gate source: dc69b4fa7c2f364af0b4792d7d1e2f9e9513279e
• Final gate SHA256: 35afe4ae98eaa80ed61726e161505caa1189b4cae0d48c5d544abac1edbb58db
• Phase 4 gate source: cd8c1733352487e156cd2941fe7e31733b42a64a
• Phase 4 gate SHA256: 7e412aebc093fc58875061fe3aa8c1c23d9e1493a800469dafce1c745acd0bc7
• Legacy retirement SHA256: beb06b4e663d88f60bc8fa343b0dec13505a3f6036c7494717d13ec54f2a83e5
• Phase 5 gate source: 8fd55434482cc597aa250c54c94a6cde68f3d43e
• Phase 5 gate SHA256: eeaf3efcbff0f8fad3211f5b3690d745378f5168c1e82fdd9c816b805d8598d2
• Phase 6 gate SHA256: 7e42e18f218d5befa5dc27da9ddc405f56d2bea794f7b563e260049ecca60da6
• Paired-seed training behavior SHA256: 13c3a37c5fe587ad1f0e6890f19791d12f4ac428ba0aaf16df59c8d59b3f51a7
• DR performance SHA256: 58dea886865aee7fb011e07bdaf7fc33aef7a33d5adb1b0391892da336bb8481
• W&B: N/A; evidence is frozen in repository artifacts and independently recomputed by validators.
• ONNX/checkpoint: exercised by the entrypoint/task-rollout gates; no release model is attached to this PR.

Checklist

• Added or updated tests where needed
• Updated docs and support matrix
• Linked the driving issue
• Noted unsupported routes and support limits explicitly
• Synced latest main with merge ancestry before final evidence
• Completed two local reviews: owner/hot-path structure and claim-to-artifact provenance
• GitHub CI passes on integration HEAD 5a8a26691e33bb3c9af1d088c89bb7a36c221d62 (run 30742715153; Docs 30742715151)
• User review/approval received

Final Remote Acceptance

• PR fix: declare CI aggregate audit dependencies #870 fixed the aggregate dependency contract by declaring pyyaml in the isolated dev group and adding fail-closed workflow/dependency tamper tests. It merged into the integration branch as 5a8a26691e33bb3c9af1d088c89bb7a36c221d62.
• Final local gate: 2352 passed, 26 skipped, 449 deselected, 1 xfailed, coverage 68%; Ruff, mypy, pyright and benchmark smoke passed.
• Final evidence: source dc69b4fa7c2f364af0b4792d7d1e2f9e9513279e, SHA256 35afe4ae98eaa80ed61726e161505caa1189b4cae0d48c5d544abac1edbb58db, 12/12 mandatory A/B/C/D runs and zero mandatory skip/xfail/xpass/deselect.
• Real aggregate ran on Python 3.12.3 with only uv sync --only-group dev, installed PyYAML 6.0.3, audited 8 shards / 227 test files / 6 local-evidence nodes, combined A-H coverage to 67%, and passed --fail-under=25.

Per #705, this PR must remain open for user review/approval. Do not squash/rebase the final merge because evidence receipts depend on preserved commit ancestry.

Issue #853 Repository Cleanup

• Fixes Chore: 清理 #847 的 issue-id 命名、入口膨胀、临时文件与 artifact 体积 #853
• 将 manager/mjwarp acceptance 的 issue-specific 命名改为语义化命名，并把产品 support evidence 移到 src/unilab/support/ + conf/support/。
• 将 acceptance 命令收敛为 scripts/audit_acceptance.py 与 scripts/capture_acceptance.py；相对 main 的顶层新增脚本总数为 3。
• 新增入口不导入、不注册 src/unilab/cli.py。
• 禁止 Git LFS：未新增 .gitattributes 或 LFS pointer，并增加 no-LFS / changed-file 1 MiB repository policy 与 CI gate。
• 删除 6 个多 MB、可再生成的 raw JSON，改用小型 receipt；原文件仍可从 Git 历史恢复。
• src/、conf/、docs/ 不再包含 issue.?705 命名。

#853 Validation

通过：

uv run ruff check .
uv run ruff format --check .
uv run mypy src/unilab
uv run pyright
uv run pytest --collect-only -q
uv run --no-sync scripts/audit_ci_test_shards.py
# targeted hygiene/support/entrypoint/repository-policy tests
git diff --check

make test-all 已执行但未通过：2185 passed, 35 failed, 14 errors, 29 skipped, 652 deselected, 1 xfailed。失败集中在已删除 raw artifact / 旧冻结 hash，以及一个单独运行可通过的顺序相关 Motrix 测试。用户于 2026-08-03 明确要求跳过这些非关键失败并尽快更新 PR，因此未重复执行 full gate；本节记录该显式 override，不将失败表述为通过。

Cleanup commit: 28c619af.

---

Phase 0 delivery (2026-08-04)

本 PR 分支历史已按 #887 重写(force-push 28c619af → 71946170,purge tests/acceptance/ 大型 raw artifacts,源码提交保持不变),顶层提交完成 #886 Phase 0 scope reset(production device-resident mjwarp 路径退役 + 显式退役诊断)。原交付 PR #890 因 head 已被本分支包含而自动关闭。

Closes #886
Closes #887

Closes #888(交付 PR: #892,合入本分支后随本 PR 合并 main 自动关闭)

[TATP-233]

History rewrite notice (#887): 本分支历史已于 2026-08-04 重写并 force-push(28c619af → 71946170)。

• 重写范围:仅本 feature branch(main..HEAD),main 与其他远程分支引用未变化;
• 清除内容:tests/acceptance/ 下全部已提交的大型 raw artifact(最大 53MB)与 issue-specific generated evidence;正常源码提交顺序与内容保持不变(80 个纯 artifact 空提交被剪除);
• 顶层新增 Phase 0: Scope reset and repository cleanup (#705 feature branch) #886 Phase 0 scope reset 提交(production device-resident mjwarp 路径退役,详见 PR refactor: Phase 0 scope reset + history rewrite for #705 feature branch (#886, #887) #890 的说明,该 PR 因 head 已被本分支包含而自动关闭);
• 原历史离线 bundle 备份保留在执行机器 /tmp/issue-705-feature-branch-backup.bundle;

协作者请重新同步本地分支:

git fetch origin
git switch feat/issue-705-manager-mjwarp
git reset --hard origin/feat/issue-705-manager-mjwarp

注意:旧提交(含 artifact)在本仓库 GitHub 对象库中可能仍经 PR ref 缓存可达,彻底 GC 需 GitHub 侧处理;新可达历史已不再包含这些对象。

[TATP-233]

#883