GeezerVison

Pathfinder.resolve(goal) returns immediately for a previously visited goal even when a new root is waiting in the frontier. Lowering an existing root's distance also leaves cached distances unchanged:

import numpy as np
import tcod.path

pf = tcod.path.Pathfinder(tcod.path.SimpleGraph(
    cost=np.ones((9, 13), np.int8), cardinal=1, diagonal=0))
pf.add_root((0, 0), value=7)
pf.resolve()
pf.add_root((0, 0), value=-3)
pf.resolve((8, 12))
print(pf.distance[8, 12])  # 27 before this change; should be 17.

When the frontier is nonempty, queue the visited goal at its current distance. The search can then find a shorter route, while the queued goal bounds the work if no improvement is possible. An empty frontier still returns immediately. Add regression coverage for C/F indexing, partial/full initial resolution, heuristics enabled/disabled and a cached-goal control that checks the search does not resolve the entire map.

On three MovingAI Dragon Age maps, 2,268 of 2,304 lazy queries retain stale distances after lowering a destination's weight. All distances match independent Dijkstra calculations with this change. In a separate integration run with #191, adding destinations produces 1,600 incorrect routes out of 2,304 before this change and none afterward. #191 addresses predecessor termination; this PR addresses when the frontier is searched. Each has a separate regression on current main.

Validation on Linux/Python 3.12 with the pinned native libtcod: 12 new cases fail before the change; all 14 pass afterward. The suite excluding window and audio tests passes (197 passed, 94 skipped); Mypy and project formatting hooks pass. Full audio-inclusive validation is incomplete: the dummy audio run stalls in stream binding, while the ordinary audio configuration lacks a device. No complete graphical/audio-suite pass is claimed.