Daikoku PA: two perimeter ramps, far side served by the median opening

Rework the PA junction to match the reference drawing: yellow = PA -> highway
(entry), red = highway -> PA (exit). Entry and exit swap sides (entry now
EAST, exit WEST) and both trunks hug the PA perimeter instead of running
concentric spirals.

The simplification: both carriageways sit at the same deck level, adjacent,
so the median opening IS the far-side access path. The under-deck crossing
and merge runs are deleted outright -- 8 runs collapse to 4, and the
girder-soffit ceiling logic that constrained the crossing goes with them.
Part count 24,210 -> 21,691.

The median opening is now load-bearing rather than cosmetic (it is the only
route from the far carriageway to the ramps), so it widens from 6 to 9
samples per junction.

Orientation is a single flip-able constant (ORIENT) so a mirrored result is
a one-line correction, not a rewrite.

Measured on the built map (not asserted):
  - 600 entry + 600 exit ramp parts
  - max grade 8.00% / 8.27%, under the 9% limit
  - 0 collisions with piers/masts/girders (558 structural parts)
  - segment continuity <= 2.5 studs; trunk->merge handoff <= 1.7 studs
  - merge lands at dY 0.0 (exact deck level), 33 studs lateral
  - median open at junctions: 0 parts vs. control ~7 on plain mainline

Two traps worth recording, both caught by controls:

1. Studio's require() cache silently rebuilt STALE code. The first "passing"
   build reported 17,575 parts and zero ramps, and the map contained
   PAOffRamp/PAOnRamp parts whose names exist nowhere in the current source.
   Every measurement taken before this was describing an old map. Added
   tools/rebuild_fresh.luau, which clones the module to bust the cache; use
   it rather than requiring MapBuilder directly.

2. A reported "median never opens" bug was a bad probe, not a defect. It
   counted `Barrier` (the outer edge parapet, which must NEVER open) instead
   of `Median_<i>`. Re-run against the right name: control 7, junctions 0.

Both were only visible because the probes carry controls -- a 0-hit result
with no control is vacuous. The verification harnesses are kept in tools/
with their controls built in.

tools/diag_raycast.luau

@@ -42,7 +42,12 @@ for i = 1, n do
 			end
 			if d > 1 then
 				hits[#hits + 1] = ("i=%d(%s) d=%.2f hitA=%s hitB=%s"):format(
-					i, samples[i].piece, d, a1.Instance.Name, a2.Instance.Name)
+					i,
+					samples[i].piece,
+					d,
+					a1.Instance.Name,
+					a2.Instance.Name
+				)
 			end
 		end
 		local b1 = workspace:Raycast(o1, Vector3.new(0, -120, 0), roadOnly)
@@ -58,7 +63,10 @@ end
 
 out[#out + 1] = ("ALL-GEOMETRY worst=%.3f at i=%d"):format(worstAll, atAll)
 out[#out + 1] = ("ROAD-ONLY  worst=%.3f at i=%d -> %s"):format(
-	worstRoad, atRoad, worstRoad <= 0.60 and "PASS" or "FAIL")
+	worstRoad,
+	atRoad,
+	worstRoad <= 0.60 and "PASS" or "FAIL"
+)
 out[#out + 1] = ("outlier samples (d>1), first 12 of %d:"):format(#hits)
 for k = 1, math.min(12, #hits) do
 	out[#out + 1] = "  " .. hits[k]