Carry plate boundaries on their own motion, closing the reconstruction

[veillette]

Running the geological-time clock tore the map open. Every plate outline was
rotated rigidly about its own Euler pole, but a plate boundary belongs to two
plates at once, so each one was carried by whichever plate PB2002 happened to
name first — driving it into its neighbour. At the ends of the slider the two
sides of a typical boundary ended up ~1600 km apart: gaps along the ridges,
overlaps at the trenches, and a Mid-Atlantic Ridge glued to Africa's margin
instead of sitting in the ocean it opens.

Boundaries now ride rotations of their own. A ridge or a transform takes the
mean of its two plates' rotation vectors, which is where a spreading axis sits
when accretion is symmetric and is stationary with respect to a fault the plates
merely slide along. A subduction zone takes the overriding plate, read from the
separator PB2002 writes into each boundary name (NZ\SA is Nazca going down
under South America, TO/PA the Pacific under Tonga). Because the rule depends
on what a boundary is rather than how it was spelled, sections named AF-AN
and AN-AF no longer ride different plates and tear a ridge in half.

Plate outlines are carried by that boundary network rather than by the plate
inside them, each vertex taking a distance-weighted blend of the boundary
motions near it. The blend is a function of position alone, so two plates
sharing an edge carry it identically and the mosaic stays a mosaic; what changes
through time is each plate's area, growing along its ridges and shrinking at its
trenches. Outlines are subdivided until an edge stops stretching, which is what
stops one drawing a straight line across an ocean where boundaries hand over.

Measured over the slider's full range, outline edges stretched by up to 6070 km
before and by at most 700 km now, with none over 1000 km; the sixteen labelled
plates hold their area to within a factor of four.

Also here:

• PlateReconstruction indexes MOTION_FRAMES — the plates first, so a plate
  index is still a frame index, then the derived boundary rotations.
• npm run build-data takes step names, so the PB2002 model can be rebuilt
  without pulling a newer earthquake catalogue and a fresh DEM into the diff.
• Two checks on the poles themselves, from data they were not derived from:
  PB2002's published relative velocities are reproduced to a median 0.04 mm/yr
  across all 1580 segments, and the no-net-rotation integral leaves a residual
  of 0.009 °/Myr.
• doc/model.md records what carries what, and is explicit that the microplates
  stop meaning anything first: ten of the fifty-two turn more than half a
  revolution over 50 Myr, and a spinning microplate drags its neighbour's edge
  with it.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01BiBoCsTcS5UQ9hewbcBEpE

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.