Feature parity audit: function & parameter level vs jsfeat (Axis 1)

[kalwalt]

Enumerate every jsfeat src/*.js symbol and map it to jsfeatNext at function-and-parameter level: signature match, default args, argument order, and numeric behavior. Produce per-module parity tables. See docs/jsfeat-parity-and-refactor-audit.md (Axis 1).

[kalwalt]

Progress: the parity audit is being executed as characterization tests in PR #49 — every function tested against the original jsfeat is a verified row of this audit (executable, not just documented).

Modules verified so far (function + parameter level)

Module	Functions	Parity	Notes
math.get_gaussian_kernel	1	✅ match (float tol, 4 cases)
matmath	12/12	✅ match	identity, transpose, multiply, multiply_ABt, multiply_AtB, multiply_AAt, multiply_AtA, identity_3x3, invert_3x3, multiply_3x3, mat3x3_determinant, determinant_3x3
transform	3/3	✅ numeric match, ⚠️ signature divergence	see findings

Findings

1. jsfeat's transform module was never shipped. It exists in src/jsfeat_transform.js on master but is absent from build/jsfeat.js and the npm package (jsfeat@0.0.8); only math.perspective_4point_transform made it into the bundle. The test oracle (tests/vendor/oracle.cjs, vendored from inspirit/jsfeat@4c7b336) evaluates the src module against the bundle to restore it.
2. Signature divergence (Axis 2): original jsfeat.transform.* functions take raw arrays (mat[0]…); jsfeatNext's transform methods take matrix_t and read .data. Same math — confirmed numerically — different calling convention.
3. The npm jsfeat@0.0.8 oracle does include haar and bbf — usable later as oracles for the porting work (Port haar module (Haar cascade object/face detection) #43, Port bbf module (BBF face detection) #44).

Next modules queued: linalg, remaining math, imgproc, structs (matrix_t/keypoint_t/pyramid_t).

[kalwalt]

Progress update (PR #49): math and linalg fully verified against the vendored jsfeat oracle — 30 tests passing.

Module	Functions	Parity
math	get_gaussian_kernel, qsort (full + sub-range), median, deprecated perspective_4point_transform	✅ match
linalg	lu_solve, cholesky_solve, svd_decompose (U_T|V_T), svd_solve, svd_invert, eigenVV	✅ match (SVD/eigen to 4 decimals)

Additional finding:

• jsfeatNext's deprecated math.perspective_4point_transform is numerically identical to the jsfeat bundle's, but adds a console.warn deprecation notice — minor behavioral divergence, documented in the test.

Next: imgproc (grayscale, blurs, pyrdown, derivatives, integral image, equalize_histogram, canny, warps, resample, hough).

[kalwalt]

Progress update (PR #49): imgproc fully verified — 44 tests passing overall.

imgproc function	Parity
grayscale (RGBA default + BGRA code), resample (u8 fast path), box_blur_gray (scaled + NOSCALE), gaussian_blur, pyrdown, scharr_derivatives, sobel_derivatives, compute_integral_image (sum+sqsum+tilted), equalize_histogram, canny, warp_perspective, warp_affine	✅ bit-for-bit match
hough_transform	⚠️ original jsfeat is broken here — see finding

Finding — original jsfeat bug, fixed in jsfeatNext: jsfeat's hough_transform references min_theta/max_theta without declaring them; under the distributed bundle's "use strict" IIFE this throws ReferenceError, i.e. the shipped jsfeat function is unusable. jsfeatNext declares them, fixing it. The test pins both behaviors (oracle throws, jsfeatNext returns sane [rho, theta] lines).

Remaining for full coverage: fast_corners, yape, yape06, orb, optical_flow_lk, motion_estimator (needs a seeded Math.random to be deterministic), structs (matrix_t ops), resample f32 path.

[kalwalt]

Progress update (PR #49): detectors, ORB, optical flow and motion_estimator verified — 52 tests passing.

Module	Parity
fast_corners (set_threshold + detect)	✅ identical corners (x, y, score)
yape06 (detect, default thresholds)	✅ identical keypoints
yape (init + detect)	✅ identical keypoints — ⚠️ API divergence: original is a static namespace, jsfeatNext requires new
orb.describe	✅ identical 32-byte descriptors
optical_flow_lk.track	✅ identical status + positions (2-level pyramid, seeded fixture)
motion_estimator.ransac + homography2d	✅ identical model + inlier mask (seeded Math.random both sides)
motion_estimator.lmeds + homography2d	✅ identical model + mask
motion_estimator.ransac + affine2d	🐛 jsfeatNext BUG — crashes (original jsfeat works) → #51

Bug found by the suite (#51): jsfeatNext's motion_model base class lost the default check_subset() (original returns true; only homography2d overrides it), so RANSAC/LMEDS with an affine2d kernel throws TypeError. Pinned in the test as a documented divergence until fixed.

Remaining (minor): matrix_t struct ops, resample f32 path.

[kalwalt]

Final batch (PR #49): structs + helpers verified — matrix_t (construction, resize grow/shrink, copy_to), get_data_type/get_channel/get_data_type_size across all popular formats (⚠️ API divergence: top-level functions in jsfeat vs instance methods in jsfeatNext), and imgproc.resample f32 path. Planned Phase 0 coverage is complete: 57 tests, all passing. Full summary in the PR #49 description. Still open for future coverage: haar/bbf oracles exist in the vendored bundle for when #43/#44 port them.