Add chatdsg_repair example + constraint-aware prompts + pddl_repair_tool

[JY-HIM4U]

Adds a natural-language interface to the plan-repair goal_manager (from the spot_tools/omniplanner plan-repair work). A user can say "go to the loading dock but avoid the bridge" instead of hand-writing PDDL goals and constraint JSON. All 6 files are new; nothing existing is modified.

Changes

examples/chatdsg_repair/chatdsg.py
Runnable entry point. Based on examples/chatdsg/chatdsg.py. Reuses its InputDisplayApp TUI via dynamic module load instead of copying it, and additionally imports pddl_repair_tool (for its registration side-effect) before the agent loads.

examples/chatdsg_repair/agent_config.yaml
Agent config. Based on examples/chatdsg/agent_config.yaml, swapping the tool from send_pddl_goal to send_pddl_goal_with_constraints, and pointing planner_topic at /hilbert/commanded_goal (goal_manager's input) instead of directly at omniplanner's region_rearrange_objects_pddl/pddl_goal. So, goals go through the pause/coverage-check/replan flow instead of bypassing it.

examples/chatdsg_repair/agent_prompt.yaml
System prompt. Based on examples/chatdsg/agent_prompt.yaml, rewritten for the goal-sending (not just querying) workflow, plus rules for when to attach forbidden-poi/forbidden-edge constraints.

examples/prompts/common/pddl_domain_description_with_constraints.yaml
Extends pddl_domain_description.yaml with a <Runtime Constraints> section documenting forbidden-poi/forbidden-edge syntax and lowercase-ID rules for the LLM.

examples/prompts/common/pddl_in_context_examples_with_constraints.yaml
Extends pddl_in_context_examples.yaml with worked examples of constrained and multi-goal calls, in the tool-calling response style this agent actually uses (the original's examples use a bare <answer> tag format that doesn't apply here).

src/heracles_agents/tools/pddl_repair_tool.py
New tool. Based on pddl_calling_tool.py's send_pddl (same robot_name/planner_topic binding pattern), extended to also accept constraints_json. A JSON list the LLM can generate naturally (e.g. [["forbidden-poi","o188"]]), which gets translated internally into the proper ConstrainedPddlGoalMsg/ConstraintFact structure before publishing (subprocess.run instead of the original's os.system).