feat(action-engine): add SeedGraph planning and bundle generation

[skywhite1024]

Description

Stack

• Stack: #543
• Layer: 4/7
• Base: ljd/gen-sim-refactor-03-task-frontend
• Depends on: feat(task-engine): add task interpretation and canonical scene binding #534
• Followed by: feat(action-engine): add runtime grounding and action adapters #536

Add deterministic and online SeedGraph planning, strict compiler and linker stages, offline/online candidate selection, bundle generation, packaged robot and sensor templates, artifact hashing, and headless graph visualization.

The persisted graph is validated through the same loader boundary used by later execution layers.

Refs #531

Type of change

• New feature (non-breaking change which adds functionality)

Validation

• Focused compiler/planning/generation coverage - 203 passed
• Stack-tip GenSim regression - 706 passed, 9 warnings
• Black 26.3.1 - 855 Python files unchanged at the stack tip
• git diff --check - passed

Checklist

• Code passes Black 26.3.1.
• Tests cover the affected behavior.
• Stack-level architecture documentation is included by layer 7.
• No new third-party dependency is required.

[greptile-apps]

Greptile Summary

The PR adds SeedGraph planning, compilation, runtime loading, visualization, and generation of complete Action Engine bundles from Prompt2Scene exports.

• Adds deterministic conversion between semantic recipes, SeedGraph v3, and runtime ExecutionPrograms.
• Adds offline and online planning, graph linking, candidate selection, and persisted runtime contracts.
• Adds scene normalization, artifact generation, CLI integration, templates, and focused test coverage.
• The fan-in conversion currently produces an inaccurate source-state topology in generated graph visualizations.

Confidence Score: 4/5

The fan-in state-topology defect should be fixed before merging because generated bundle diagrams can misrepresent the execution graph.

Multi-parent execution units preserve all dependency IDs but use only one parent state as their source, while visualization derives its topology solely from those source-to-target transitions.

Files Needing Attention: embodichain/gen_sim/action_engine/compiler/v2.py and embodichain/gen_sim/action_engine/graph_visualization.py

Important Files Changed

Filename	Overview
embodichain/gen_sim/action_engine/compiler/v2.py	Adds bidirectional SeedGraph conversion, but fan-in units materialize an inconsistent source-state topology that produces inaccurate graph visualizations.
embodichain/gen_sim/action_engine/generation/generator.py	Orchestrates task interpretation, planning, scene normalization, validation, and bundle generation with extensive input checks.
embodichain/gen_sim/action_engine/generation/source_scene.py	Resolves Prompt2Scene formats and normalizes identifiers, paths, scales, poses, and world-frame rotation.
embodichain/gen_sim/action_engine/planning/linker.py	Links node and task-group dependencies while validating persisted execution contracts and resource constraints.
embodichain/gen_sim/action_engine/runtime/loader.py	Loads and verifies persisted SeedGraphs, supports regeneration and A/B artifacts, and converts them into runtime programs.
embodichain/gen_sim/action_engine/graph_visualization.py	Adds chain and DAG rendering with runtime overlays; its source-target topology exposes the compiler's fan-in representation defect.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Prompt2Scene export] --> B[Scene preparation]
    B --> C[Task planning]
    C --> D[SeedGraph linking and validation]
    D --> E[SeedGraph artifact]
    D --> F[ExecutionProgram conversion]
    F --> G[Runtime loader]
    F --> H[Graph visualization PNG]
    B --> I[Fast gym config]
    E --> J[Generated bundle]
    H --> J
    I --> J

Loading

Prompt To Fix All With AI

### Issue 1
embodichain/gen_sim/action_engine/compiler/v2.py:238
**Fan-in state topology drops branches**

When an execution unit has multiple parent units, this conversion uses only the first parent’s target as the edge source while retaining every parent in `depends_on`. Because graph visualization derives topology exclusively from `source` and `target`, generated bundle diagrams disconnect the remaining branches from the join and can assign incorrect DAG levels and layout.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "feat(action-engine): add SeedGraph plann..." | Re-trigger Greptile}

[greptile-apps]

Fan-in state topology drops branches

When an execution unit has multiple parent units, this conversion uses only the first parent’s target as the edge source while retaining every parent in depends_on. Because graph visualization derives topology exclusively from source and target, generated bundle diagrams disconnect the remaining branches from the join and can assign incorrect DAG levels and layout.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/action_engine/compiler/v2.py
Line: 238

Comment:
**Fan-in state topology drops branches**

When an execution unit has multiple parent units, this conversion uses only the first parent’s target as the edge source while retaining every parent in `depends_on`. Because graph visualization derives topology exclusively from `source` and `target`, generated bundle diagrams disconnect the remaining branches from the join and can assign incorrect DAG levels and layout.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.