feat(evaluation): add agent-confirmed PointNav evaluation

[Staaaaaaaaar]

Motivation

Implement roadmap PR 09 with a deterministic PointNav evaluator, while correcting two architecture gaps discovered during implementation:

• navigation completion must be declared by the Agent instead of inferred from a transient goal crossing;
• concrete Task/Evaluator selection needs one composition root in the Experiment container.

Scope

• Add the AgentTaskState ROS contract with IDLE, RUNNING, SUCCEEDED, and FAILED states.
• Require an Agent to stop producing non-zero commands before reporting SUCCEEDED.
• Implement ordered 3D trajectory sampling and simple PointNav metrics in simulation time.
• Confirm Agent success against the first ground-truth pose at or after the Agent result.
• Add evaluator integration through the orchestrator's serialized termination path.
• Detect a frozen simulation clock with a steady-time watchdog.
• Add rh_bringup as the sole concrete Experiment-container composition root.
• Remove Task, Evaluator, and generic lifecycle application entrypoints.

Out of scope

• Result persistence and Experiment summaries.
• Multi-Episode execution and retry policy.
• Collision metrics, SPL, PoseNav, and automatic goal sampling.
• Isaac Sim, GPU, and production three-container deployment validation.

Architecture changes

The final success decision now uses two signals:

1. the Agent stops its navigation control and reports SUCCEEDED;
2. the Evaluator confirms that ground-truth position remains inside the configured PointNav radius.

Passing through the goal without Agent completion no longer terminates the Episode. Agent failure, simulation-time timeout, infrastructure failure, and operator abort remain distinct termination paths.

rh_experiment remains implementation-independent and exposes Task/Evaluator factory boundaries. rh_bringup is the only package that imports concrete implementations and currently provides the reviewed pointnav_simple assembly.

Implementation

• Add C++/Python typesupport and contract coverage for AgentTaskState.
• Extend the Mock Agent with transient-local task-state publication and sequence checks.
• Add pure trajectory and evaluation modules plus a TF2-backed ROS observer.
• Add evaluator submission and clock-watchdog support to the single-Episode orchestrator.
• Add ros2 run rh_bringup experiment --profile pointnav_simple.
• Update repository lint discovery and architecture documentation.

Testing

• Clean development-container build using new temporary build/install/log directories: 16 packages built.
• Clean full test run: 192 tests, 0 errors, 0 failures, 0 skipped.
• Repository structure, Markdown, YAML, Ruff, and git diff --check: passed.
• Clean install entrypoint audit: only rh_bringup experiment is exposed for the central runtime.
• Integration coverage verifies TF transformation, fixed trajectory metrics, goal pass-through without termination, Agent completion plus ground-truth confirmation, no evaluator control publisher, and frozen /clock handling.

Acceptance criteria

• Scope and out-of-scope match the implementation.
• Build, lint, and relevant tests pass.
• Documentation is updated.
• Acceptance evidence is included.

Known limitations

• Metrics remain in memory until the Recorder PR.
• The current orchestrator still intentionally supports one Episode.
• Agent task state is an initial compact protocol; richer feedback and failure taxonomy remain extensible.
• CPU mock integration is not evidence of Isaac Sim or robot-controller compatibility.

Follow-up

• PR 10: durable result recorder.
• PR 11: multi-Episode execution and failure policy.
• Runtime deployment work will assign the bringup command to the Experiment service in the three-container stack.