pub const MAX_FIT_STATIONS: usize = 1024;Expand description
Absolute station ceiling, whatever the caller’s budget.
The ladder doubles the interpolation budget from what the caller asked for while the MEASUREMENT says the curve misses its tolerance, and stops here — or at the input’s own station count, whichever comes first.
1024 is a COST cap, not a geometric one. As SHIPPED it is a BACKSTOP that
nothing in the corpus reaches, because [LADDER_BUDGET_FACTOR] binds first
for every caller here (80 → 160 for the imprint driver, 96 → 192 for
offset_shell). The reading below is the ladder let OFF that bound, kept
because it is what the ceiling exists for. Measured over the 55-row case
population: 282 fits, 21 of which miss at the caller’s own budget; 16 of
those reach their tolerance inside the ladder, and 5 do not — three at this
ceiling with 3,070 control points and two stalled at 2,302 — every one of
them an offset_shell section on the LOCAL interpolation lane, which
converges at h² and cannot reach 1e-7 on a curved section at any station
count this input offers. The lane is the fix there, not the ceiling; raising
the ceiling buys accuracy in the fourth significant figure and pays for it
in control points on every consumer downstream. A fit that stops here
reports PolylineFitExit::StationCeiling with the deviation it actually
holds, the way a pcurve fit reports its sample ceiling.