Terrain rewrote every junction's scoring economics, so the tuner re-ran (12
varied launches per event, closed-loop aim, percentile targets). The spread is
now $3k bronze at Cliffs Corner to $111k gold at Albert x Adelaide -- terrain
funnels that one into a monster pile-up, while the downhill scenery-crash
events (Margaret x Edward, Cliffs Corner, Bridge Approach) top out early with
median == max around $3-4k: winnable and honestly tuned, but they end on the
first wall. Those three still deserve hand-picked crossings eventually.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Medal targets are no longer one uniform 15/35/60k guess across every event.
tools/tune_junctions.gd plays each junction 12 times with varied launches and
sets bronze/silver/gold at the 35th/65th/88th percentile of actual damage, so
they now range from $4k at Cliffs Corner to $73k gold at Melbourne x Merivale.
The old numbers made half the events impossible and the other half trivial.
Grey x Glenelg scored $0 in 12/12 runs and took three fixes, each found by
measuring rather than reading:
- The launch sat on the street centreline, so the car drove cleanly BETWEEN the
two convoy lanes -- closest pass 2.9 m, almost exactly the 2.7 m lane offset.
It now launches in a lane. (The first attempt offset the wrong way and widened
the pass to 5.5 m, which is how the sign error announced itself.)
- int(L / headway) floors to one car on a short clipped stream, so that junction
fielded 9 convoy cars against the usual 20-25. Minimum 3 per stream, evenly
spaced to fit.
Result: $0 -> $91k, convoy 9 -> 18, and 0 of 9 junctions now unplayable.
The tuner has turned out to be the event validator as much as a balance tool --
"scores nothing in every run" is an unplayable event, not a tuning problem.
Also from the review: the deformer no longer crumples det_* panels after they've
been shed onto their own body, and notes that rebuilding an ArrayMesh drops any
imported LODs.
Honest gap: Cliffs Corner and Bridge Approach show median == max at ~$4.7k --
winnable and correctly tuned, but the run ends on scenery before the convoy is
reached. Both are curved approaches; they want better crossings picked by hand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
9 junction events across 6 levels, auto-generated from real OSM intersections
(tools/osm_junction.py): each crossing street's ways become lane-offset convoy
streams, deterministic speeds/intervals/models -- learnable puzzles, not dice.
Menu lists CRASH JCT entries; crash -> 9s settle with camera-relative
aftertouch on the wreck and a one-shot crashbreaker shockwave (blast() on
traffic, radial impulse on debris, orange boom). Damage tally vs
bronze/silver/gold targets. Fixed zero-normal UV projection that striped
roads. Smoke test drives a full junction run to tally (k, gold).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
