feat(viewer): report which faces touch which, and steady the zoom

[pclauss123]

Summary

Three viewer changes the DFM app needs before its directions/mapping work can land
in toolpath-template. They are separated from that work deliberately — this is the
package half, and it is the half that has to publish first.

regionAdjacency(model, geometry) — new export. Reads the mesh and returns, for
each region, the regions sharing an edge with it. That is what lets a feature be
drawn by clicking one face and following the surface, instead of picking every face
by hand. PartMesh takes an onAdjacency callback and computes it once per mesh
rather than per query.

Steadier zoom.Viewer gains zoomTo: 'cursor' | 'centre', and re-frames the
part on a double middle click (recentreOnDoubleClick) — the way back from having
zoomed into a corner, which zooming to the cursor makes easy to do.

On the middle button because double left click is where a viewer usually puts "orbit
about this from now on". That gesture is not in this PR: two implementations of it
were tried against a real part and neither took effect in the app, and rather than
guess at the cause a third time it is left for a later pass. Putting re-framing on the
left button meanwhile would only have to move again.

The middle gesture is paired from single presses by trackDoubleTaps, also exported,
because dblclick fires for the primary button only — there is no such event for the
middle button however many times it is pressed.

DirectionArrows's shownDirection widens to accept a list as well as a single
index, so more than one way up can be shown at once.

Two behaviour changes worth reading, neither an API break

A part rebuilt on the same geometry used to blank the new one. A consumer whose
report changes identity — a feature added, a re-fetch — rebuilds the part against the
same cached mesh, and React builds the new part during render before disposing the
old. The old part's dispose deleted the region attribute unconditionally, including
the one the new part had just set, so every vertex fell back to texel 0: the whole
part in one flat colour, with hover, selection and every wash gone with it. dispose
now removes the attribute only if it is still the one that part set. Covered by
part.test.ts — "leaves a newer part on the same geometry alone".

Five of the nine DIRECTION_COLORS are retuned so neighbouring directions stay
apart when the part is washed by direction. Same export, same length, same type — but
anything hard-coding a hex or screenshot-testing the part will see a difference.
Called out in the changeset for that reason.

Why minor

Checked name by name, because it decides the bump. Every symbol the consuming app
imports is already in 0.3.1 except regionAdjacency. The surface change is one
new export, three new optional props, one widened prop type — no removals, no
signature changes, no peer-dep changes.

Note main already carries an unreleased squared-view-cube changeset, also minor,
so 0.4.0 will contain both this and the view cube fix. Rebased onto that work;
squaredUp and regionAdjacency coexist in the built entry point.

Public package release

• This PR changes no public package source.
• I ran pnpm changeset and committed the generated Changeset.
• This source change intentionally needs no release; I applied the
  no-release-needed label and explained why below.

Validation

• I ran the relevant checks and tests.

pnpm --filter @toolpath/viewer test 19 files, 212 tests, green
pnpm --filter @toolpath/viewer check-types clean
pnpm --filter @toolpath/viewer build clean, dist/ emitted
node scripts/check-release-intent.mjs origin/main
Release intent recorded for: @toolpath/viewer

Confirmed regionAdjacency, zoomTo, recentreOnDoubleClick and the widened
shownDirection all reach dist/index.d.ts — the consuming app type-checks against
the published types, not against source, so that mattered to check.

[pclauss123]

Updated after running it against a real part — worth a re-read rather than a re-skim, since two things changed substantively.

The render/part.ts story in the original description was wrong. I had it as a RedFormat/unpackAlignment crash fix. It is not: the file still builds an RGBAFormat texture and unpackAlignment appears nowhere. The real change is a dispose guard — a part rebuilt on the same cached geometry had its region attribute deleted by the old part's dispose, so every vertex fell back to texel 0 and the part went one flat colour. part.test.ts covers it. The description and the changeset now say so, and a dead RedFormat import is gone.

The double-click gestures were rebound. Double left click used to re-frame. It now brings the clicked point to the middle and orbits about it from then on; re-framing moved to a double middle click. This matches visualization/part.tsx in the current app. Worth knowing that setOrbitPoint was the obvious primitive and is deliberately not used — it moves the pivot without moving the camera, so the gesture looked like it did nothing.

214 tests, types and build clean.

[pclauss123]

One more update, and this one makes the PR smaller.

The double-left-click orbit-centre gesture I mentioned in the last comment is out. Two implementations were tried against a real part — setOrbitPoint, then re-targeting the camera the way visualization/part.tsx does — and neither took effect in the app. Whatever is swallowing it is not in the handler, and a third guess is not how to find it, so it is left for a later pass rather than shipped half-working. The handler, the transient marker and the carriedPosition helper all came out with it.

Double middle click still re-frames the part, which does work, and that is the only gesture this PR adds. It stays on the middle button because the left one is where the orbit-centre gesture belongs when it lands.

Net effect on the review: one fewer public prop (pivotOnDoubleClick is gone), ~180 fewer lines in viewer.tsx, and render/camera.ts is untouched again. 212 tests, types and build clean.