refactor(motion_estimator): de-duplicate motion_estimator and extract kernels (#47)

[kalwalt]

Eighth slice of #47 — the biggest structural win of the series.

What changed

• New src/motion_model/motion_model.ts — the motion_model base plus the affine2d and homography2d kernels, verbatim from the monolith (they were module-local classes there; in original jsfeat they live under the jsfeat.motion_model namespace). Kernels instantiate linalg via direct module import (established pattern).
• src/motion_estimator/motion_estimator.ts — type-only stub replaced by the real implementation (get_subset, find_inliers, ransac, lmeds); lmeds instantiates math directly.
• src/core/core.ts — the temporary any-typed affine2d/homography2d static slots (a documented TODO since refactor(math): extract core base class and de-duplicate math module (#47) #62) now have precise typeof types. No any slots remain in core.
• src/jsfeatNext.ts — down to ~380 lines (from ~3,900): pure aggregator except for optical_flow_lk, the last inline module.

Verification

• tsc --noEmit → clean (it caught two missing imports mid-extraction; fixed before commit)
• npm test → 57/57 — three parity tests pin ransac/lmeds producing identical models and inlier masks vs original jsfeat on seeded data (incl. the affine2d kernel fixed in fix(motion_estimator): restore affine2d error() and check_subset() (#51) #52)
• UMD bundle smoke-checked: instanceof for the estimator and both kernels; affine RANSAC recovers the synthetic ground-truth model
• Public API unchanged: jsfeatNext.motion_estimator, .affine2d, .homography2d, .ransac_params_t all attach exactly as before (motion_model remains unexposed, as it always was)
• dist//types/ untouched (rebuilt when the integration branch merges to dev)

Advances #47. One module left: optical_flow_lk.

🤖 Generated with Claude Code