sharpen 04 vs 09 differentiation (keep both)

GPT Pro's curriculum review suggested merging 04_online_replanning_astar
into 09_blocked_path_recovery. After re-reading both examples I disagree:
the trigger for replanning is genuinely different.

04 plans optimistically through UNKNOWN cells and replans when the lidar
*observes* a hidden wall. The replan trigger is passive observation; no
action ever fails.

09 plans on a fully known map. A dynamic blocker appears mid-run. The
agent tries to step into that cell, the world returns
Failure(kind='blocked_path'), and the agent has to recover (step back,
mark the cell) before replanning.

These are two distinct lessons: observation-driven belief revision
versus execution-failure recovery. Both stay. The docstrings and the
navigation README now state the contrast explicitly, and the gap audit
records the keep-both decision.

Author: rsasaki0109

docs/implementation_gap_audit.md

@@ -28,6 +28,7 @@ continuous-control Gymnasium adapter, and the curiosity exploration example.
 | `09_active_viewpoint_for_grasp.py` | Viewpoint selection by expected reliability under pose belief. |
 | `21_object_permanence_toy.py` | Memory persists across an occluder. |
 | `22_where_did_i_see_it.py` (removed) | The "explore -> memorize -> query -> revisit" loop overlapped `21_object_permanence_toy.py` without adding a different lesson; removed after a GPT Pro curriculum review. |
+| `04_online_replanning_astar.py` (kept) | GPT Pro suggested merging into 09. After review, kept because the trigger for replanning is *passive observation* of an unknown wall, while 09's trigger is *execution failure*. Both docstrings and READMEs were updated to make the distinction explicit. |
 | `23_model_error_recovery.py` | Regime shift detection then short system-identification probe. |
 | `24_information_gain_navigation.py` | Active scout to reveal a gate before A* with full info. |
 | `25_clear_path_before_pick.py` | Precondition failure recovery by clearing an obstacle. |

examples/navigation/04_online_replanning_astar.py

@@ -1,4 +1,19 @@
-"""Online A* replanning when new obstacles are observed."""
+"""Online A* replanning when new obstacles are observed.
+
+The agent plans through UNKNOWN cells under the optimistic assumption that
+they are free. When a hidden wall enters the lidar field of view it is
+revealed as OCCUPIED, the path becomes invalid, and A* runs again on the
+updated map.
+
+Success: robot reaches the goal cell.
+Failure: timeout (terminal).
+
+Compare to `09_blocked_path_recovery.py`. That example triggers replanning
+through an *execution failure* (the agent tried to step into a blocked
+cell and got back a `Failure`). This example triggers replanning through
+*passive observation* (the lidar reveals a wall before the agent ever
+touches it).
+"""
 
 from __future__ import annotations
 

examples/navigation/09_blocked_path_recovery.py

@@ -1,4 +1,21 @@
-"""Detect a blocked path, recover, and replan around it."""
+"""Detect a blocked path, recover, and replan around it.
+
+The map is fully known. A dynamic blocker appears mid-run on the planned
+path. When the agent tries to step into that cell the world returns
+`Failure(kind="blocked_path", recoverable=True)`. The agent steps back,
+marks the cell, and replans on a map that treats the blocked cell as
+occupied.
+
+Success: robot reaches the goal cell.
+Failure: blocked_path (recoverable), collision (recoverable), timeout
+(terminal).
+
+Compare to `04_online_replanning_astar.py`. That example triggers
+replanning through *passive observation* (lidar reveals an unknown wall).
+This example triggers replanning through *execution failure* (the agent
+attempts a step, the world rejects it, and the agent has to recover
+before replanning).
+"""
 
 from __future__ import annotations
 

examples/navigation/README.md

@@ -114,8 +114,15 @@ observe moving obstacle -> predict next cell -> choose safe move -> observe agai
 ### What this teaches
 
 A plan is only valid relative to the robot's current map. The agent plans
-through unknown cells, observes a hidden wall, invalidates the path, and runs
-A* again.
+optimistically through unknown cells, observes a hidden wall through the
+lidar, invalidates the path, and runs A* again. The trigger for replanning
+is *passive observation*, not action failure.
+
+Success: robot reaches the goal cell.
+Failure: timeout (terminal).
+
+Compare to `09_blocked_path_recovery.py`, where the trigger is execution
+failure rather than observation.
 
 ### Run
 
@@ -287,7 +294,16 @@ observe obstacle -> roll out candidate controls -> choose lowest cost -> act ->
 ### What this teaches
 
 Failure is part of the API. A path can become blocked during execution, and the
-robot should detect that failure, update memory, recover, and replan.
+robot should detect that failure, update memory, recover, and replan. The
+trigger for replanning is *execution failure* (`Failure(kind="blocked_path")`),
+not observation.
+
+Success: robot reaches the goal cell.
+Failure: blocked_path (recoverable), collision (recoverable), timeout
+(terminal).
+
+Compare to `04_online_replanning_astar.py`, where the trigger is observation
+of a previously unknown wall before the agent ever touches it.
 
 ### Run