test: edge cases, data structures and buffer-pool coverage (refs #87, categories B + D)

[kalwalt]

Category B of the #87 plan (edge/boundary inputs). 19 new cases, 180 → 199 total (198 passing + 1 expected failure). Tests only, no src/ changes.

Follows #106 (linalg + matmath), #108 (imgproc), #109 (math + detectors + optical flow) which completed category A.

Why

Degenerate inputs are where CV code breaks, and the rest of the suite runs on comfortable 16×16 and 96×72 images that say nothing about them. This covers 1×1, single rows and columns, odd-vs-even dimensions, kernels larger than the image, all-black/all-white, borders wider than the image, 1×1 and singular matrices, and minimum kernel sizes.

It found a real bug — #114

box_blur_gray does not preserve a uniform image when the image is smaller than the kernel. A blur is an average, so a constant image must round-trip at any size. It doesn't:

image	radius	output
1×1	1	[85, 85]	wrong
4×4	2	[51, 128]	60% low
5×5	2	[128, 128]	correct

The rule is exact — output settles precisely at size >= 2*radius + 1, verified for radius 1–4. A separate float-truncation defect makes radius 3 come out 1 low even on a 32×32 image, because 128 * 49 * (1/49) = 127.999999999999986 truncates to 127.

Both are bit-identical in original jsfeat, so they are inherited rather than jsfeatNext regressions, and gaussian_blur has neither — it is correct from 1×1 upward. This is precisely the class of defect parity testing structurally cannot find, since the oracle shares it.

A follow-up posted on #114: the wrong output is not deterministic. The identical 1×1 radius-1 call returns 85 in isolation and 142 when run after the other cases in the same file, because the window reads outside the image and picks up unrelated state.

How the bug is pinned

With it.fails, not by asserting the wrong numbers — those aren't stable, per the above. The body asserts the correct behaviour and the test passes only because that assertion currently fails, so fixing #114 flips it to a failure and forces a deliberate update.

Verified that guard is not a no-op: making the inner assertion trivially true turns the suite red.

What else is covered

Everything below confirms an invariant still holds at the boundary:

• gaussian_blur preserves a constant at 1×1, 1×8, 8×1, 2×2, 3×3, 5×4, 4×5 for kernels 3/5/7
• sobel and scharr derivatives stay exactly zero; canny finds nothing
• equalize_histogram maps a constant to a single value rather than inventing structure
• integral-image corner equals the total, first row and column zero, at every degenerate shape
• all-black and all-white survive a blur unchanged, derivatives zero
• detectors return 0 rather than throwing on tiny images and on borders wider than the image
• lu_solve reports failure on a singular system and still solves a well-conditioned one (so the 0 is a meaningful signal, not "always fails")
• svd_invert handles 1×1; size-1 gaussian kernel is the identity; even-size kernels stay normalized and symmetric

Plus one more characterization: invert_3x3 on a singular matrix silently returns NaN/±Infinity with no signal to the caller — the same family as #102.

Verification

prettier clean · tsc --noEmit clean · license-check clean · npm test 198 passed + 1 expected fail.

Remaining #87 scope

Category A and B are now done. Still open: C third-party ground-truth fixtures, D coverage gaps (data_type has no dedicated test). As discussed, C is the natural point to revisit the empirically-characterised bounds in the earlier phases with tighter assertions.

🤖 Generated with Claude Code

[kalwalt]

Pushed category D (coverage gaps) onto this branch as well, since it was still unmerged and both parts are tests-only. Retitled accordingly. 34 further cases, 199 → 233 total.

Prompted by the question "is #87 actually complete?" — an audit of what the suite touched found more holes than the issue listed:

Component	Before
node_utils/data_t	zero references anywhere
node_utils/_pool_node_t	zero
point_t	zero
pyramid_t, keypoint_t, ransac_params_t	constructed as scaffolding, never asserted on
cache (shared pool)	api-shape only: typeof get_buffer === "function"

The buffer pool is the important one. AGENTS.md makes "balance every get_buffer with a put_buffer" a standing rule, every algorithm borrows from this one pool, and an imbalance stays invisible until the pool drains and an unrelated module starts reading someone else's scratch memory. There is now a test holding every module to that rule across 23 operations — imgproc, the detectors, orb, optical_flow_lk, the linalg solvers. All balanced today.

Mutation-checked, since a leak detector that cannot detect leaks is worse than none: deleting one put_buffer in box_blur_gray and another in fast_corners.detect produces

AssertionError: expected 'box_blur_gray: 29' to be 'box_blur_gray: 30'
AssertionError: expected 'fast_corners.detect: 28' to be 'fast_corners.detect: 29'

— caught, and the message names the offending operation.

Also covered: data_t's 8-byte alignment (the f64 view would throw without it) and its four views aliasing one buffer; matrix_t's view selection, zero-fill, and the reuse-vs-reallocate rule in resize; pyramid_t level sizing and that a uniform image stays uniform at every level; ransac_params_t.update_iters respecting its cap and rising with the outlier ratio.

Two things I checked rather than assumed:

1. point_t is not on the namespace — but neither is jsfeat's, and nothing in src/ ever constructs one (four modules import it purely as a type annotation). So the absence is parity, not a gap. The test now records that instead of asserting a constructor that shouldn't exist.
2. data_type was already covered indirectly via the core wrappers in tests/parity/structs. The new cases pin all sixteen type/channel combinations and tie the reported element size to the buffer matrix_t actually allocates.

With this, #87 categories A, B and D are complete. Only C (third-party ground-truth fixtures) remains.

Verified: prettier clean, tsc --noEmit clean, license-check clean, npm test 232 passed + 1 expected fail.