fix(brain): destroy retired skill instances' ROS entities on reload (navigator leak)

[theo-michel]

Problem

Every skill reload leaks ROS entities. Two compounding paths, neither with teardown:

1. Double instantiation per load — discovery calls SkillLoader._get_name(), which builds a throwaway instance of each skill class just to read .name; the catalog then builds the kept one. For navigate_to_position each instance is 3 BasicNavigator nodes (~10 action clients each). The sim skill's throwaway adds another navigator even in real-robot mode (it's only dropped after discovery, in _apply_sim_swap).
2. No retirement on reload — reload_all() / reload_selective() overwrite the catalog dicts, orphaning old instances. Orphans are cyclic garbage (Node → subscription → bound callback → Node), so their zenoh graph entities stay registered until a gen-2 GC pass — which an idle process may not run for hours.

Measured on the live robot (mars-the-44th, 2026-07-07)

• 8 duplicate navigator entity sets (~90 subscriptions) alive in the skills server across /, /mapfree, /navigation after an 83-reload storm (a client hammered /brain/reload at ~100 Hz for 30 min — separate bug, tracked separately).
• One triggered /brain/reload_primitives: +2 entity sets in /mapfree & /navigation, +3 in root (sim throwaway), +19.5 MB RSS, 39.5 s duration.
• Isolated repro: orphans accumulate across reload cycles; a single gc.collect() releases all of them — nothing hard-pins them, they're just uncollected cyclic garbage.

Fix

• Skill.shutdown() lifecycle hook (no-op default). Deliberately leaves shared-server-node entities alone — destroying entities under a spinning executor is unsafe (fix(brain): stop destroying ROS entities under spinning executors (skills-server cancel crash-loop) #497).
• Nav2Controller.destroy() / SimPathPlanningController.destroy() — destroy the navigator nodes, including assisted_teleop_client, which Humble's BasicNavigator.destroy_node() misses and whose live handle keeps the rcl node registered (destroy_when_not_in_use).
• SkillLoader._get_name — shuts the throwaway instance down.
• SkillRepository — retires replaced/pruned instances on reload_all, reload_selective, and _prune_stale_skills.
• SkillsActionServer — disposal gate: retired instances are destroyed immediately when idle, or deferred to execute_callback's finally while a skill runs. A mid-run reload can retire the currently executing instance, whose entities its execute() is still spinning; destroying those live would recreate the fix(brain): stop destroying ROS entities under spinning executors (skills-server cancel crash-loop) #497 crash class. Navigator nodes are never attached to any executor when no skill is executing, so idle-time destruction is exactly as safe as the GC destruction that already happens today — just deterministic.

Verification

• With gc.disable(), three simulated reload cycles against the live graph hold at exactly 1 entity set (previously +2/cycle) and drop to 0 on final retire, including /assisted_teleop/_action/* and the sim skill's nodes.
• Fast no-ROS pytest bucket (test_backwards_compat.py, 12 loader/discovery tests) passes with the patched loader — the new shutdown() call in _get_name is exercised by every discovery test.
• py_compile clean on all six files; 120-col ruff limit respected.

Note for deploy: brain_client needs a rebuild + skills-server restart; the two workspace/innate_skills/ files hot-reload on their own afterward.

[greptile-apps]

Greptile Summary

This PR adds deterministic cleanup for skill reloads. The main changes are:

• Adds a Skill.shutdown() lifecycle hook.
• Retires replaced skill instances during reloads and pruning.
• Defers retired-instance disposal while a skill is running.
• Destroys navigator-owned ROS nodes during shutdown.
• Cleans up temporary discovery instances after reading skill names.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

Important Files Changed

Filename	Overview
workspace/innate_skills/navigate_to_position.py	Adds shutdown cleanup for the real navigation controller and its navigator nodes.
workspace/innate_skills/navigate_to_position_sim.py	Adds shutdown cleanup for the sim navigator and path-planning node.
ros2_ws/src/brain/brain_client/brain_client/nodes/skills_server.py	Adds the disposal gate for retired skill instances during skill execution.
ros2_ws/src/brain/brain_client/brain_client/skills/catalog.py	Retires replaced and pruned code-skill instances during reload operations.
ros2_ws/src/brain/brain_client/brain_client/skills/loader.py	Shuts down temporary skill instances created during name discovery.
ros2_ws/src/brain/brain_client/brain_client/skills/types.py	Adds the base shutdown hook for skill-owned resource cleanup.

_{Reviews (3): Last reviewed commit: "fix: don't let a failed assisted_teleop ..." | Re-trigger Greptile}