pointToSegmentClosestPoint loses finite projections at extreme coordinate scales

pointToSegmentClosestPoint can return NaN or the wrong endpoint for finite, nonzero segments when the squared length overflows or underflows. This is independent of the final distance-norm issue in #51: the closest-point API does not compute a distance.

Reproduced on main 7b26935a5cccb6b68a3d889d5e5d03ba04dc9a1b (v0.0.38), Bun 1.4.2. For a horizontal segment (0,0) -> (4*s,0) and query (2*s,3*s), the exact closest point is (2*s,0):

Reversing the small segment changes the wrong answer to x/s=4. All input coordinates and the expected result are finite and representable. These are extreme-scale robustness cases, not a claim that ordinary PCB coordinates fail.

Runnable regression (tests/projection-scales-repro.test.ts, bun test tests/projection-scales-repro.test.ts):

import { expect, test } from "bun:test"
import { pointToSegmentClosestPoint } from "../src"

for (const scale of [1, 1e160, 1e-170]) {
  for (const reverse of [false, true]) {
    test(`closest point at ${scale}, reverse=${reverse}`, () => {
      const left = { x: 0, y: 0 }
      const right = { x: 4 * scale, y: 0 }
      const a = reverse ? right : left
      const b = reverse ? left : right
      for (const [x, expectedX] of [[2, 2], [-1, 0], [5, 4]]) {
        const actual = pointToSegmentClosestPoint(
          { x: x * scale, y: 3 * scale }, a, b,
        )
        expect(actual.x / scale).toBe(expectedX)
        expect(actual.y).toBe(0)
      }
    })
  }
}
test("zero-length segment remains a point", () => {
  const endpoint = { x: 2, y: 3 }
  expect(pointToSegmentClosestPoint({ x: 4, y: 7 }, endpoint, endpoint)).toEqual(endpoint)
})

Observed: 3 tests pass, 4 fail; TypeScript check passes. In src/segment-distance.ts, dx_ab * dx_ab + dy_ab * dy_ab becomes Infinity or zero. The former makes the interior projection Infinity/Infinity; the latter incorrectly selects the degenerate-segment branch. A scale-aware projection should retain interior projection and endpoint clamping without treating a nonzero segment as a point. Other helpers repeat the squared-length formula and may need the same treatment; this reproduction isolates the closest-point API. #44 focuses on distance performance rather than this function.

Prepared with Codex (Astra) assistance and verified locally.

Start in src/segment-distance.ts at pointToSegmentClosestPoint and run bun test tests/projection-scales-repro.test.ts. Check the reported extreme-scale cases, including reversed and zero-length segments. Done means the regression passes, finite nonzero segments retain the expected closest point, and the zero-length case remains unchanged.