The Model Predictive Path Integral (MPPI) con troller shipped with ROS 2 Nav2 is a widely used deployable local controller for mobile robots, yet in narrow, crowded, dynamic spaces it can stall in local minima, and its critics avoid collisions only through soft costs, with no formal safety guarantee. Methods exist for each weakness, and several works outside Nav2 address both together, but we know of no public Nav2 controller plugin that combines local-minima escape with a barrier-based safety filter. This system paper presents Safe-Escape MPPI (SE-MPPI), a Nav2 controller plugin that integrates online local-minima detection-and-escape for MPPI with a control-barrier-function (CBF) safety filter for tracked dynamic obstacles. A first version of the plugin reached the goal in none of its 44 valid live trials on a correct build; we report the four defects that we identify as the cause (one inferred, not logged) and the repaired controller. Its escape carries four additions to the advised design: two of them, avoid lists and a movement watchdog, were needed for the development sweeps of a family of pockets to pass (ablated jointly; the ablation is confounded with position in the sweep), and two are safeguards for other configurations. We disclose all four, with a joint ablation of the first two. In a pre-registered live campaign (one post hoc amendment to its crash-stop rule, after which all s4 SE-MPPI trials ran), SE-MPPI without recoveries reached the goal in 5 of 6 on s1 (Wilson 95% interval [0.44,0.97]) and 5 of 6 on the held-out s4, which lies outside the development family; stock without recoveries in 0 of 6 on each; with Nav2’s default recovery tree, on s1, stock 0 of 6 and SE-MPPI 3 of 6, on one simulated host (Fisher exact p = 0.0152 for each of the first two contrasts and p = 0.1818 for the third, which is descriptive). s1 is the development scenario; s4 carries the claim. A third pocket, s3, that has a near mirror twin in the development family gave 4 of 6 against 0 of 6 (p = 0.0606, not significant). We make no claim beyond one simulated host, six repetitions per cell, and no real robot. In a randomized, paired 1,200-trial benchmark on a Python 2D mirror that transcribes the shipped algorithm, not on the Nav2 controller itself, escape raises success on the static U-trap family from 0% to 80% (Wilson 95% interval [67,89]); a second arm of 1,200 trials on the same 200 seeded worlds, with the four additions switched off, reaches 58%. The escape–safety coordination matches an independent escape+CBF stack there, a null result, so we claim it only as a mechanism. By a standard argument for time-varying CBF gains, the gain schedule preserves forward invariance of the barrier’s safe set, a conditional guarantee that covers only a look-ahead point and the tracked dynamic obstacles, assumes that the barrier inequality holds between control updates, and requires a slack free quadratic program.