feat(brain): add provisional continuous visual navigation

[axelpey]

Summary

• keep nav_insight_continuous(target_description) as a built-in mode in the new local-agent loop
• add innate-os/continuous_navigation as a separately launchable, long-running skill with its own camera and Gemini decision loop
• reassess fresh camera frames while Nav2 is still moving; keep an unchanged waypoint without restarting the base, or cancel and replace an in-flight waypoint when the visual route changes
• ground model-selected floor pixels into stable odom-frame goals so model and cancellation latency do not stale the next command
• stop active motion on reached, blocked, user cancel, deactivation, navigation failure, or external interruption

Two implementations

1. Local agent mode: the agent retains the objective and supervises bounded go_to_point_in_view -> navigate_to_position actions from its normal turn loop.
2. Standalone skill: continuous_navigation owns the full loop inside one skill run and can be launched directly from the skill surface.

Both use Nav2 for path planning and obstacle avoidance. Neither changes the configured drive-speed limits. Replans are only issued for a materially different visual goal so repeated reassessment does not continually reset acceleration.

Verification

• focused local-brain plus standalone-skill tests: 67 passed
• standalone skill appears live as innate-os/continuous_navigation with no load error
• live Low Poly Town run: initial decision, in-motion reassessment, objective reached at the traffic lights
• ruff check and ruff format --check on all affected files
• git diff --check

This remains a draft because the standalone visual policy is intentionally a direct local port and still needs physical-robot validation.