Add Evo Studio cloud training, VLA-RL backend, dataset pool, and account ledger

[pearlq12345]

What this adds

This PR wires RoboClaw into a full embodied AI experiment platform: from dataset upload and curation rewards, through AI-assisted training planning, to cloud job orchestration and auto-repair — all driven by a single natural language intent.

---

Cloud training orchestration

roboclaw/cloud/evo_train.py — TCP bridge to EVO_Train backend. Handles training start/stop/status, checkpoint resolution, SSH runtime binding, and stage translation.

roboclaw/http/routes/train_cloud.py — Full cloud training lifecycle API: plan, start, stop, status, billing settlement, runtime matching, source preflight, supervisor repair, SSH bind.

roboclaw/training/ — Training application layer:

• service.py — orchestrates local vs cloud routing
• schema.py — TrainingStartSpec, TrainingJobStatus, TrainingPlanSpec
• vla_rl.py — VLA/RL plan validation and deployability checks
• ai_planner.py — LLM-driven intent → structured training plan
• rlinf_catalog.py — discovers RLinf configs, infers algorithm/benchmark/model metadata, sets RLINF_EXT_MODULE for worker processes
• agent_consult.py — deterministic backend agent for Evo Studio consult surface

Supervisor auto-repair: when a cloud job fails, the supervisor watches the job, infers the failure cause, and retries on the same runtime under configurable automation policies (ask / safe_auto / full_auto).

---

VLA-RL backend

roboclaw_vla/rl/ — launcher, model registry, adapters, evaluator. Supports --suite=<name> for LIBERO benchmark selection without needing to know internal Hydra config names.

roboclaw_vla/config/rl/ — LIBERO-10 GRPO config with pi0 model and FSDP training backend.

roboclaw/embodied/policy/ — policy registry with ACT, Diffusion, GR00T, pi0, SmolVLA.

---

Dataset pool and curation rewards

roboclaw/data/dataset_adapters/ — LeRobot adapter with format mapping and registry.

roboclaw/data/ingestion.py — materializes HuggingFace, local path, and archive sources into the local catalog.

Datasets uploaded and shared through quality checks earn contributor reward points tracked in the account ledger.

---

Account and billing

roboclaw/account/ledger.py — credit ledger with training holds, topup orders, reward points, and atomic hold reassignment.

roboclaw/account/training_billing.py — hourly cost estimation and service fee calculation.

---

Agent tools

roboclaw/agent/tools/cloud_training.py — EvoStudioCloudTrainTool: balance check, backend probe, runtime match, source preflight, job start/stop/status, wallet integration.

roboclaw/agent/tools/evo_studio_agent.py — EvoStudioAgentConsultTool: natural language → cloud control actions via the consult surface.

Both tools are only registered when embodied_service is available, so CLI-only agent sessions are unaffected.

---

Frontend

• Training Center: CloudIntentPanel, CloudSourcePanel, CloudProviderPanel, CloudPlanResultPanel
• Training progress: failure detection, auto-repair entry, loss curve, stage timeline, automation mode selector
• Dataset page: upload, ingest, storage usage, credit display

---

Bug fixes

• scan.py: ROBOCLAW_DISABLE_CAMERA_SCAN env var for test isolation
• loop.py: deduplicate _CLOUD_PROBE_EXEC_PATTERNS against base.py
• runtime.py: remove silent exception swallowing in provider model init
• evo_train.py: validate that workflow or dataset_name is provided before calling bridge
• loop.py: guard Evo Studio tool registration behind embodied_service is not None
• data/curation/validators.py: fix circular import between dataset_adapters and curation

---

Configuration

Cloud training requires a running EVO_Train backend. See .env.example for all required and optional environment variables (ROBOCLAW_EVO_TRAIN_HOST, ROBOCLAW_EVO_TRAIN_PROVIDER, etc.).

RLinf-based training recipes require cloning https://github.com/RLinf/RLinf.git and setting ROBOCLAW_RLINF_REPO_PATH. Without it, the catalog returns empty and logs a warning — local and cloud training without RLinf still work.

---

Tests

598 passing, 234 skipped. Frontend builds clean.

New test files: test_evo_train_routes, test_vla_rl_routes, test_account_ledger, test_dataset_adapters, test_curation_reward, test_policy_registry, test_launcher_import, test_dataset_upload_completion, test_command_builder_train