feat(task-engine): add task interpretation and canonical scene binding

[skywhite1024]

Description

Stack

• Stack: #543
• Layer: 4/8
• Base: ljd/gen-sim-refactor-02-seedgraph-contracts
• Depends on: feat(task-engine): add TaskSpec and semantic task-graph contracts #533
• Followed by: feat(action-engine): add SeedGraph planning and bundle generation #535

Add the strict Task Engine frontend that turns structured instruction candidates into provider-free task intent and binds language roles once to canonical scene identities.

Architecture boundary

• Adds strict TaskDraft, SceneRequirements, semantic call candidate, and bound-draft values.
• Consumes the JSON-safe projection emitted by ExpertProgramIntegrationCatalog.planner_projection() and requires its exact integration fingerprint.
• Decodes deterministic and structured-model outputs through the same local TaskSpec and canonical semantic-call validators.
• Supports multiple independent candidates, candidate-local failures, deterministic deduplication, and model provenance.
• Resolves roles only through canonical SceneManifest references (SceneObjectRef, SceneArticulationRef, SceneLinkRef, and SceneAffordanceRef). Unknown, ambiguous, or mistyped identities fail explicitly.
• Keeps language-role binding separate from physical goal grounding; no poses, qpos, trajectories, ActionInvocations, controllers, solvers, or live providers are produced.
• Removes robot/action config construction, packaged arm profiles, hard-coded arm fields, and the parallel Action Engine scene inventory.

Refs #531

Type of change

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

Validation

• pytest -q tests/gen_sim/task_engine - 48 passed
• pytest -q tests/docs/test_check_api_docs.py - 8 passed
• Black 26.3.1 - 789 Python files unchanged
• Public API docs - 1862/1862 exports documented
• git diff --check - passed
• Full Sphinx build not run locally because sphinx-build is unavailable in the validation environment

Checklist

• Code passes Black 26.3.1.
• Tests cover canonical codec reuse, catalog fingerprint matching, typed SceneManifest binding, explicit failure, and shared model/deterministic validation.
• Public API documentation is aligned.
• No new third-party dependency is required.

[greptile-apps]

Greptile Summary

The PR adds scene-independent semantic task interpretation, candidate generation, scene grounding, TaskSpec assembly, scene handoff validation, and simulator configuration generation.

• Introduces strict Task Engine drafts, scene requests, success specifications, and candidate-set contracts.
• Adds LLM-backed instruction interpretation and scene-reference grounding for E1–E9 tasks.
• Adds Action Engine TaskSpec lowering and scene handoff validation.
• Adds robot-profile-driven simulator and agent configuration builders.

Confidence Score: 4/5

The PR is not safe to merge until the simulator templates required by build_fast_gym_config are added or its template references are corrected.

The semantic interpretation and grounding contracts appear internally consistent, but every invocation of the new gym-config builder reaches missing packaged templates and raises FileNotFoundError before producing a configuration.

Files Needing Attention: embodichain/gen_sim/action_engine/generation/config_builder.py and embodichain/gen_sim/action_engine/generation/templates/

Important Files Changed

Filename	Overview
embodichain/gen_sim/task_engine/interpretation.py	Adds strict LLM instruction interpretation, normalization, transport configuration, and intent validation.
embodichain/gen_sim/task_engine/agent.py	Adds concurrent candidate generation, canonicalization, deduplication, and vote aggregation.
embodichain/gen_sim/task_engine/contracts.py	Defines and validates JSON-safe task drafts, scene requests, success specifications, and candidate sets.
embodichain/gen_sim/action_engine/tasks/grounding.py	Adds fail-closed semantic scene-reference grounding with cardinality and eligibility checks.
embodichain/gen_sim/action_engine/tasks/interpretation.py	Bridges validated Task Engine intents into grounded Action Engine TaskSpecs.
embodichain/gen_sim/action_engine/tasks/assembly.py	Adds scene inventory modeling and E1–E9 TaskSpec and requirement assembly.
embodichain/gen_sim/action_engine/generation/config_builder.py	Adds simulator and agent manifest generation, but its required robot, sensor, and light templates are absent, making the new gym-config API fail immediately.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
I[Task instruction] --> TI[Task Engine interpretation]
TI --> D[Validated semantic draft]
D --> G[Scene-reference grounding]
S[Scene inventory] --> G
G --> TS[Grounded TaskSpec and scene requirements]
TS --> H[Scene handoff validation]
H --> C[Simulator and agent config generation]
C --> R[Action Engine runtime]

Loading

Prompt To Fix All With AI

### Issue 1
embodichain/gen_sim/action_engine/generation/config_builder.py:225
**Required configuration templates are missing**
When `build_fast_gym_config()` runs, `_make_robot()` and the subsequent sensor and light loaders reference template files that are absent from this directory, causing generation to raise `FileNotFoundError` before returning a gym configuration. A/B mode also requires the absent `vlm_sensors.json`; please include the referenced templates or correct these references.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "feat(task-engine): add semantic task int..." | Re-trigger Greptile}

[greptile-apps]

Required configuration templates are missing

When build_fast_gym_config() runs, _make_robot() and the subsequent sensor and light loaders reference template files that are absent from this directory, causing generation to raise FileNotFoundError before returning a gym configuration. A/B mode also requires the absent vlm_sensors.json; please include the referenced templates or correct these references.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/action_engine/generation/config_builder.py
Line: 225
Comment:
**Required configuration templates are missing**
When `build_fast_gym_config()` runs, `_make_robot()` and the subsequent sensor and light loaders reference template files that are absent from this directory, causing generation to raise `FileNotFoundError` before returning a gym configuration. A/B mode also requires the absent `vlm_sensors.json`; please include the referenced templates or correct these references.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.