A single noisy DEM cell could fabricate a huge fake slope and make an edge
unconditionally impassable, forcing the pathfinder to route around passable terrain.
Replace the hard cliff (|grade|>max_grade -> skip) with a smooth exponential penalty:
no penalty up to max_grade, then base_time *= exp(overshoot * SLOPE_PENALTY_SCALE);
only grades whose penalty exceeds SLOPE_PENALTY_CAP (true bad data / vertical) are
dropped. Routing can now see through noisy cells while still strongly avoiding real
cliffs. Penalty only raises edge cost, so the heuristic stays admissible.
2 tests: smooth penalty traverses a >max_grade gap (finite, raised cost) yet routes
around / drops a past-cap grade; exactly-at-threshold grade incurs no penalty.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace MCP_Geometric in _pathfind_wilderness with a numba-jit anisotropic A* (new
astar.py): signed-slope speed (climbing != descending; tobler peaks at -0.05), hard
cliff, per-edge avg context multiplier, trail-takes-both via 256-entry lookup, per-edge
barriers (strict/pragmatic/emergency), multi-goal A* (first popped wins) with admissible
distance/base-speed heuristic. New compute_cost_multiplier_grid (slope-free context
multiplier) + exponential inflation (sigma=1.8; inf->HARD=50*p95 for blur, inf re-imposed).
numba>=0.59 added (numba 0.65.1).
fix: wilderness leg is always foot effort; mode parameter reserved for future flexibility.
_pathfind_wilderness keeps the mode param (threaded from _route_A/B/C) but hardcodes
cost_mode=foot for the cost grid, trail friction, speed function, base speed, and max
grade. Off-trail math for MTB/ATV/vehicle is not well-grounded and real-world wilderness
traversal is foot regardless (push the bike, walk past where the vehicle stops). User mode
still drives entry-point eligibility (query_radius highway filter) and Valhalla network
costing. Matches the original pre-#17 design: wilderness ALWAYS uses foot.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
